ArticlePDF Available

A review of machine learning for automated planning


Abstract and Figures

Recent discoveries in automated planning are broadening the scope of planners, from toy problems to real applications. However, applying automated planners to real-world problems is far from simple. On the one hand, the definition of accurate action models for planning is still a bottleneck. On the other hand, off-the-shelf planners fail to scale-up and to provide good solutions in many domains. In these problematic domains, planners can exploit domain-specific control knowledge to improve their performance in terms of both speed and quality of the solutions. However, manual definition of control knowledge is quite difficult. This paper reviews recent techniques in machine learning for the automatic definition of planning knowledge. It has been organized according to the target of the learning process: automatic definition of planning action models and automatic definition of planning control knowledge. In addition, the paper reviews the advances in the related field of reinforcement learning.
Content may be subject to copyright.
The Knowledge Engineering Review, Vol. 00:0, 1–24. c
2009, Cambridge University Press
DOI: 10.1017/S000000000000000 Printed in the United Kingdom
A Review of Machine Learning for Automated Plan-
Departamento de Inform´
atica, Universidad Carlos III de Madrid.
Avda. de la Universidad, 30. Legan´
es (Madrid). Spain
Recent discoveries in automated planning are broadening the scope of planners, from toy problems to real
applications. However, applying automated planners to real-world problems is far from simple. On the
one hand, the definition of accurate action models for planning is still a bottleneck. On the other hand,
off-the-shelf planners fail to scale up and to provide good solutions in many domains. In these problematic
domains, planners can exploit domain-specific control knowledge to improve their performance in terms
of both speed and quality of the solutions. However, manual definition of control knowledge is quite
difficult. This paper reviews recent techniques in machine learning for the automatic definition of planning
knowledge. It has been organized according to the target of the learning process: automatic definition of
planning action models and automatic definition of planning control knowledge. In addition, the paper
reviews the advances in the related field of reinforcement learning.
1 Introduction
Automated Planning (AP) is the branch of Artificial Intelligence that studies the computational synthesis
of ordered sets of actions that perform a given task. AP appeared in the late ’50s as the result of studies
into state-space search, theorem proving and control theory to solve the practical needs of robotics
and automatic deduction. The STanford Institute Problem Solver STRIPS (Fikes and Nilsson, 1971),
developed to be the planning component for controlling the autonomous robot Shakey (Nilsson, 1984),
perfectly illustrates the interaction of these influences. From the days of Shakey to now, AP has produced
accepted standards for representing planning tasks and efficient algorithms for solving them (Ghallab
et al., 2004). In the last decade AP systems have been successfully applied to real world problems
such as planning space missions (Nayak et al., 1999), managing fire extinctions (Castillo et al., 2006)
or controlling underwater vehicles (Bellingham and Rajan, 2007).
Despite these successful examples, the application of AP systems to real world problems is still
complicated, mainly because of two knowledge definition problems:
Automated planners require an accurate description of the planning task. These descriptions include
a model of the actions that can be carried out in the environment, a specification of the state of the
environment and the goals to achieve. In the real world, the execution of an action may result in
numerous outcomes, the knowledge of the state of the environment may be partial and the goals may
not be completely defined. Generating exact definitions of the planning tasks in advance is unfeasible
for most real-world problems.
Off-the-shelf planners often fail to scale-up or yield good quality solutions. Usually, the search for a
solution plan in AP is a PSpace-complete problem (Bylander, 1991, 1994). Current state-of-the-art
planners try to cope with this complexity through reachability analysis by (1) grounding actions and
(2) performing a search process guided by domain-independent heuristics. Nevertheless, when the
number of objects is large, the resulting ground search trees cannot be traversed in a reasonable time.
Moreover, where heuristics are poorly informed –like in the case of some domains with strong sub-
goals interactions– this kind of analysis is misleading. Domain-specific search control knowledge
has been shown to improve the scalability of planners in these situations (Bacchus and Kabanza,
2000; Nau et al., 2003). Defining search control knowledge is usually more difficult than defining
the planning task because it requires expertise, not only in the task to solve, but also in the planning
Since the beginning of research in AP, Machine Learning (ML) has been a useful tool to overcome
these two knowledge acquisition problems. A comprehensive survey of AP systems that benefit from ML
can be found at (Zimmerman and Kambhampati, 2003). At the present time, there is renewed interest in
using ML to improve planning. In 2005 the first International Competition on Knowledge Engineering
for Planning Systems (ICKEPS) was held and – in 2008– a track for learning-based planners was opened
within the International Planning Competition (IPC). Furthermore, workshops on planning and learning
have periodically taken place within the International Conference on Automated Planning and Scheduling
(ICAPS). Even if this review includes classic systems for planning and learning, mainly it focuses on
covering the last approaches on ML for AP.
The review is organized according to AP’s two knowledge acquisition problems: (1) learning planning
action models and (2) learning search control. Figure 1 illustrates these two targets of the ML process.
Traditionally, most planners separate the definition of the action models from the definition of search
control knowledge to allow the planner the use of different representations. There are cases, however,
where this division between models and control knowledge is unclear like, for example, macro actions
that can be placed in both categories. The paper also reviews the latest techniques in Reinforcement
Learning (RL) (Kaelbling et al., 1996; Sutton and Barto, 1998) which are very closely related to the ML
for AP task. In RL the agent interacts with its environment to automatically learn an action-value function
of the environment. This action-value function combines information from both the search control and
action model to solve a particular task.
For each of the reviewed techniques, the paper provides a unified study analyzing its scope, and
strengths and weaknesses, with respect to four common issues in the ML process: knowledge representa-
tion,extraction of the experience,learning algorithm and exploitation of learned knowledge. The paper
is structured as follows: the second section presents the two areas of the study, Automated planning and
Machine Learning; the third section reviews the diverse techniques for the automatic learning of planning
action models; the fourth section reviews the different approaches for automatically learning planning
search control knowledge; the fifth section describes the techniques for RL and the last section outlines
promising new avenues of research in learning for planning systems.
2 Background
This section provides an overview of the area under study, Automated Planning, and presents our criteria
for reviewing the planning and learning systems.
A Review of Machine Learning for Automated Planning 3
Figure 1 Use of ML to improve the knowledge of the AP processes.
2.1 Automated Planning
AP aims to produce solvers that are effective when faced with different classes of problems. AP produces
solvers that exploit environment dynamics models to reason on different tasks in different environments.
An AP task is defined by two elements:
The domain, that consists of the set of states of the environment Stogether with the set of actions A
that indicates the transitions between these states.
The problem, that consists of the set of facts s0Sthat indicates the initial state of the environment
and the set of facts GSthat indicates the goals of the AP task.
A solution to an AP task is a sequence of actions a1,...,an. This sequence of actions corresponds to a
sequence of state transitions s0,...,snsuch that aiis applicable in state si1and produces state si,snis
a goal state; that is Gsn. The cost of the solution is the sum of the action costs. A solution is considered
optimal when it minimizes the expression Pn
i=1 cost(ai).
Specifying accurate action models to address AP tasks in the real world is a complex exercise. Even in
traditionally easy-to-code planning domains– like Blocksworld– it is hard to specify the actions’ potential
outcomes when the environment is non-deterministic. In extreme cases, like planning for the autonomous
control of a Mars Rover (Bresina et al., 2005) or an underwater vehicle (McGann et al., 2008), this
knowledge cannot be specified because it is as of yet unknown. In these situations, the success of the AP
systems fully depends on the skills of the experts who define the action model. With the aim of assisting
experts, the planning community is developing systems for the acquisition, validation and maintenance of
AP models. ML seems to be a useful tool for these systems to automate the extraction and organization of
knowledge from examples of plans, execution traces or user preferences.
When the AP task is accurately defined, it looks like it should be easy to tackle by searching for a path in
a state-transition graph, a well-studied problem. However, in AP this state-transition graph often becomes
so large that search is intractable. The complexity of the algorithms to solve this type of problems grows
with the number of states, which is exponential in the number of problem variables (number of objects
and predicates of the problem). Prior to the mid ’90s, planners were unable to synthesize plans with more
than ten actions in an acceptable amount of time.
In the late 90’s, a significant scale-up in planning took place due to the appearance of the reachability
planning graph (Blum and Furst, 1995). This discovery allowed for the development of powerful domain-
independent heuristics (Hoffmann and Nebel, 2001a; Bonet and Geffner, 2001) that were computable
in polynomial time. Recent discoveries– such as the automatic extraction of search landmarks (Porteous
and Sebastia, 2004) and the automatic construction of symbolic Pattern Data Bases (Edelkamp, 2002)–
improve planners’ speed and quality performance. Planners using these advances can often synthesize one-
hundred action plans in seconds, yet scalability limitations are still present in AP and even well-studied
domains –like the Blocksworld– become challenging for planners when there is a relatively large number
of objects. On the one hand, current domain-independent heuristics are expensive to compute. This effect
is more evident in domains where heuristics are misleading. In these domains planners spend most of
their planning time computing useless node evaluations. On the other hand, given that these domain-
independent techniques are based on action grounding, the planners’ search trees become intractable when
the number of problem objects and/or action parameters reaches a certain size. These problems make
domain-independent planners’ application to various real problems difficult. Logistics applications need
to handle hundreds of objects as the same time as hundreds of vehicles and locations (Florez et al., 2010)
which makes computing evaluation functions in every search node unfeasible. In this case, ML has a role
to play in capturing useful control knowledge skipped by the domain-independent techniques.
2.2 Machine Learning
According to (Mitchell, 1997), there is a set of issues in common when using ML to improve an automatic
process. The review of the planning and learning systems completed in this paper is carried out in relation
to them:
1. Knowledge representation. First, the type of knowledge that the ML process will learn must be
defined. In this paper we consider two different targets of ML for AP, action models to feed planners
and search control to guide the planners’ search for solutions. Second, how to represent the learned
knowledge must be decided. In this case, two representation decisions must be made:
(a) Representation language. The kind of notation used to encode the target concept and the
experience. Because AP tasks are normally described in predicate logic this is the most
often used representation language for encoding AP concepts. That said, others languages like
description logic or temporal logic are also used, although to a lesser extent.
(b) Feature space. The set of features that the ML algorithm considers to learn the target concept. In
AP, these features are normally the domain predicates used to define the the AP task’s actions,
states and goals.
2. Extraction of experience. How learning examples are collected. In the case of AP, learning examples
can be autonomously collected by the planning system or provided by an external agent, such as a
human expert. Implementing a mechanism to autonomously collect learning examples is a complex
process. Using a planner to collect experience is an open issue mainly because guaranteeing an AP
problem’s solvability with a given domain model is just as hard as the original AP task. Random
explorations frequently under-sample AP tasks’ state and action spaces. AP actions typically present
preconditions that can only be satisfied by specific sequences of actions which have low probability
of being chosen by chance.
3. Learning algorithm. How to capture patterns from the collected experience. Different approaches can
extract these patterns. Inductive Learning builds patterns by generalizing from observed examples.
Analytical Learning uses prior knowledge and deductive reasoning to build patterns that explain
the information from the learning examples. Hybrid Inductive-Analytical Learning combines the
two previous learning techniques to obtain the benefits of both: a better accuracy in generalization
when prior knowledge is available and use of observed learning data to overcome the shortcomings
in the prior knowledge. When designing learning algorithms for AP inductive learning is the most
commonly used technique, but analytical and hybrid approaches have also used based on the domain
definition of the AP task to build explanations of the collected learning examples.
4. Exploitation of learned knowledge. How the automatic system benefits from the learned knowledge.
The decisions made on each one of the previous three issues will affect the learned knowledge. If
the learned knowledge is imperfect, it must be applied via a mechanism that guarantees a robust
exploitation. In the case of AP, the imperfect knowledge could be the result of several circumstances:
some representation choices may not be expressive enough to capture relevant knowledge for a given
A Review of Machine Learning for Automated Planning 5
domain; the strategy for collecting the learning experience may miss significant examples of the target
knowledge; or the learning algorithm may get stuck in local minima or not be able to capture patterns
of the target knowledge within reasonable time and memory requirements. In these situations, a direct
use of the learned knowledge may damage the planning process. Planning and learning systems need
to be equipped with mechanisms that allow them to plan as robustly as possible despite flaws in the
learned knowledge.
3 Learning Planning Action Models
AP algorithms reason about correct and complete action models that indicate the state-transitions of
the world. Building planning action models from scratch is difficult and time-consuming, even for AP
experts. An alternative approach is to use ML so people do not have to hand-code the action models.
This section reviews ML techniques for the automatic definition of action models for AP. The review
classifies the techniques by the stochasticity of the actions effects and by the observability of the state of
the environment.
1. Action effects. In many planning tasks deterministic world dynamics cannot be assumed. This is the
case with planning domains that include stochastic procedures such as the toggling of a coin or the
rolling of a dice, or non-deterministic outcomes, such as robot navigation in the real world.
2. State observability. In many planning tasks handling a complete and exact description of the state
of the environment is inconceivable. Parts of the current state may be confused or missing due to
sensors’ failures or to their inability to completely sense the world. For example, when controlling a
robot in the real world.
Accordingly, we defined four categories for AP modelling: deterministic actions in fully observable
environments; deterministic actions in partially observable environments; stochastic actions in fully
observable environments; and stochastic actions in partially observable environments. Despite the fact
that other classifications are possible– like for instance by grouping according to the the learning
target (preconditions, effects, conditions of effects, probabilities of outcomes, .. . )– we believe this
one is useful for planning purposes because each class corresponds to a different planning paradigm.
Figure 2 summarizes the classification of the planning action modelling systems by giving some example
implementations. This table does not intend to be an exhaustive enumeration so the systems in the table
are just a sample. The rest of the section describes the systems belonging to each of the four categories in
Strengths Weakness
Deterministic effects Learning complexity is Poor expressiveness LIV E (Shen and Simon, 1989),
Full state observability theoretically bounded EXP O (Gil, 1992),
Efficient planning algorithms OBS ERVER (Wang, 1994)
Complete coverage of the
learning examples
Deterministic effects Complete coverage of the Poor expressiveness AR MS (Yang et al., 2007),
Partial state observability learning examples Inefficient planning algorithms (Amir and Chang, 2008),
ao et al., 2008),
LOC M (Cresswell et al., 2009)
Probabilistic effects Rich expressiveness Non-existent online learning (Oates and Cohen, 1996),
Full state observability Efficient planning algorithms TRA IL (Benson, 1997),
LOP E (Garc´
ınez and Borrajo, 2000),
(Pasula et al., 2007),
PELA (Jim´
enez et al., 2008)
Probabilistic effects Rich Expressiveness High planning and (Yoon and Kambhampati, 2007)
Partial state observability learning complexity
Figure 2 Implementations of systems for planning action modelling.
3.1 Learning Deterministic Models in Fully Observable Environments
This is the simplest scenario and corresponds to the classical planning task, where planning actions present
deterministic effects and observations of the states are correct and complete. To offer an example of the
modeled actions, Figure 3.1 shows the application of the unstack(B,A) action from the Blocksworld
in a fully observable and deterministic environment. The Blocksworld is a classic domain in AP which
consists of a set of blocks, a table and a robot hand: the Blocks can be on top of other blocks or on the
table; a block that has nothing on it is clear; and the robot hand can hold one block or be empty. The
unstack(B,A) action makes the robot hand unstack block Bfrom block A.
State s_1
State s_0
Figure 3 Example of application of the unstack action in a fully observable and deterministic Blocksworld.
3.1.1 The Learning Task
1. Knowledge representation. In terms of AP, the suitable languages for representing deterministic
actions are the languages defined for classical planning. The Planning Domain Definition Language
(PD DL) (Fox and Long, 2003; Thi´
ebaux et al., 2005; Gerevini et al., 2009b) is the standard
representation language for classical planning. It was created as the planning input language for
the International Planning Competition (IPC1) to standardize the planning representation languages
and facilitating comparative analysis of diverse planning systems. Although several of the systems
reviewed in this paper use other representation languages, we have used PDDL to illustrate our
examples. A PDDL action aconsists of a tuple <head(a),pre(a),eff(a)>where,
head(a) is the head of action acontaining the action name and the list of typed parameters,
pre(a) is the set of action preconditions. This set represents the facts that needs to be true for the
application of action aand,
eff(a) is the set of action effects. This set includes the facts that are no longer true after the
application of action a(known as del effects and represented by del(a)) and the facts made true
by the application of action a(known as add effects and represented by add(a)). In PDDL, the
del effects are preceded by the logical connective not.
Figure 4 shows the PDDL representation for action unstack from the Blocksworld as an example.
(:action unstack
:parameters (?top - block ?bottom - block)
:precondition (and (emptyhand) (clear ?top) (on top? ?bottom))
:effect (and (not (emptyhand)) (not (clear ?top)) (not (on (?top ?bottom))
(holding ?top) (clear ?bottom)))
Figure 4 P DD L representation for action unstack from the Blocksworld.
2. Learning examples. Learning examples consist of tuples < si1, ai, si>where aiis the action
applied at state si1, i.e., the pre-state and siis the state produced by applying the action, i.e., the
3. Learning algorithm. When actions present deterministic effects and the environment states are fully
observable, the dynamics of a given action are captured from the lifted literals of the pre-state and
the post-state of action executions. Learning the action preconditions is accomplished by lifting the
literals from a set of pre-states. Learning the effects of a given action is completed by lifting the set
A Review of Machine Learning for Automated Planning 7
of literals that is different in the pre-states and post-states of action executions. This set of literals
is called the delta-state and is formally defined as: ∆(si1, si)=(si1\si)S(si\si1). The first
term of the delta-state captures the del effects of action aiand the second term captures its add effects.
In both learning steps– learning the action preconditions and learning the action effects– the lifting
process substitutes each grounded object with a variable. The type for each variable is given by the
most specific type in the type hierarchy of the substituted object.
4. Exploitation of the learned knowledge. In this scenario, the learned models may present unnec-
essary preconditions or missing effects. These model flaws can be eliminated by observing more
executions of the action but collecting a good observation sample of planning actions is still an open
issue. Random applications of AP actions may leave actions unexplored. Normally AP actions present
preconditions that can only be satisfied by specific sequences of actions which have a low probability
of being chosen by chance. A planner may also fail to provide solutions with intermediate learned
models. Considering any literal of the pre-state of an action as a real precondition of that action leads
to very restrictive models that fail to solve problems known to be solvable. An alternative approach is
to rely on an external source of experience that collects these extra observations of actions, a teacher
for instance.
3.1.2 Implementations
The LI VE system (Shen and Simon, 1989) was an extension of the General Problem Solver (GPS)
framework (Ernst and Newell, 1969) with a learning component. LIV E alternated problem solving with
model learning to automatically define operators. The decision about when to alternate depended on
surprises, that is situations where an action effects violated its predicted model. EXPO (Gil, 1992)
generated plans with the PRODIG Y system (Minton, 1988), monitored the plans execution, detected
differences in the predicted and the observed states and constructed a set of specific hypotheses to fix those
differences. Then the EXPO filtered the hypotheses heuristically. OBSERVER (Wang, 1994) learned
operators by monitoring expert agents and applying the version spaces algorithm (Mitchell, 1997) to the
observations. When the system already had an operator representation, the preconditions were updated by
removing facts that were not present in the new observation’s pre-state; the effects were augmented by
adding facts that were in the observation’s delta-state.
All of these early works were based on direct liftings of the observed states. They also benefit from
experience beyond simple interaction with the environment such as exploratory plans or external teachers,
but none provided a theoretical justification for this second source of knowledge. The work recently
reported in (Walsh and Littman, 2008) succeeds in bounding the number of interactions the learner must
complete to learn the preconditions and effects of a STRI PS action model. This work shows that learning
STRIPS operators from pure interaction with the environment, can require an exponential number of
samples, but that limiting the size of the precondition lists enable sample-efficient learning (polynomial in
the number of actions and predicates of the domain). The work also proves that efficient learning is also
possible without this limit if an agent has access to an external teacher that can provide solution traces on
In the review we focused on learning STRIPS-like action models but others systems have tried to learn
more expressive action models for deterministic planning in fully observable environments. Examples
would include the learning of conditional costs for AP actions (Jess Lanchas and Borrajo, 2007) or the
learning of conditional effects with quantifiers (Zhuo et al., 2008).
3.2 Learning Deterministic Models in Partially Observable Environments
In this scenario actions also present deterministic effects, but the state is not fully observable. As
a consequence, the current state is described by the set of possible states called belief state. This
scenario corresponds to the conformant planning task, when the current state is non observable, and
to the contingent planning task when the current state can be partially observed. As an example of the
actions modeled in this scenario, Figure 3.2 shows the application of the unstack(B,A) action in a
deterministic and non-observable Blocksworld. In the example, there are two blocks, A and B. The current
state is uncertain and is described by a belief state b0that represents the disjunction of five possible states.
The result of applying action unstack(B,A) is a belief state b1which represents the disjunction of four
possible states. State 1 in b0generates state 3 in b1, while the action applied to the rest of states in b0has
no effect, so they are still probable in b1.
1 2
Belief State b_0
1 2
Belief State b_1
Figure 5 Example of application of the unstack(B,A) action in a non-observable two-blocks Blocksworld.
3.2.1 The learning task
1. Knowledge representation. Because in this scenario actions are deterministic, a classical planning
representation language, such as PD DL, provides a suitable representation. Figure 4 shows the
PDDL representation for action unstack from the Blocksworld.
2. Learning examples. Learning examples consist of tuples < bi1, ai, bi>where bi1is a belief state
corresponding to the partial observation of the pre-state,aiis the action applied at the pre-state and
biis the belief state corresponding to the partial observation of the post-state.
3. Learning algorithm. In this scenario, lifting the pre-state and delta-state of observations does
not define the preconditions and effects of an action. Given that the states of the environment are
not completely known, a given observation of a pre-state cannot be associated with an action’s
preconditions. Likewise, changes of a given set of literals cannot be directly associated to the
application of a specific action. In this scenario two approaches can be used:
Inference. Inference can be used to obtain action models that satisfy the collected learning
examples. This task is closely related to the Boolean satisfiability (SAT) problem which
determines if the variables of a given Boolean formula can be assigned in such a way as to
make the formula evaluate as true. The action modeling task in this scenario can be defined as
the satisfaction of a logic formula for every action in the domain. This formula is the disjunction
of the possible conjunctions of pre-states and post-states observed for the action. Figure 6 shows
an example of the formula for obtaining models for the action unstack(X,Y) that satisfy the
observation in Figure 3.2.
Induction. One can search the space of possible action models to find the best fit for the observed
learning examples. This task is closely related to the Inductive Logic Programming (ILP) task
that searches the space of the Logic Programs2to find the best fit for a set of observed facts.
Action modeling in this scenario can be defined as the induction of two logic programs for every
action in the domain. Figure 7 shows the two logic programs that model action unstack from
the Blocksworld.
The first logic program captures the preconditions of the action. The head of this one-clause
logic program consists of the action name and parameters, and the body is the subset of literals
from the observed pre-states that best fits the learning examples. This learning process can be
2Here, the concept of Logic program is used in a narrower sense generally applied, i.e., as a conjunction of disjunctions
of literals where each disjunction presents at most one positive literal. This type of disjunction is called a Horn clause.
A Review of Machine Learning for Automated Planning 9
(or ;;; pre-state 1 post-state 1
;;; pre-state 1 post-state 2
(and (Unstack B A)
(DelClear B) (DelOn B A) (DelOntable A)
(AddClear A) (AddOn A B) (AddOntable B))
;;; pre-state 1 post-state 3
(and (Unstack B A)
(DelEmptyhand) (DelClear B) (DelOn B A)
(AddHolding B) (AddClear A))
;;; pre-state 5 post-state 5
(and (Unstack B A))
Figure 6 Example of a logic formulae for modeling action unstack from a two-blocks Blocksworld.
preconditions_unstack(TOP,BOTTOM):- object(TOP),object(BOTTOM),
emptyhand(), clear(TOP), on(TOP,BOTTOM).
effects_unstack(TOP,BOTTOM):- object(TOP),object(BOTTOM),
del_emptyhand(),del_clear(TOP), del_on(TOP,BOTTOM),
Figure 7 Representation of action unstack from the Blocksworld as two logic programs.
implemented as a heuristic search guided by the number of observed pre-states of the action
covered and not covered by the different potential programs. The second program captures the
effects of the actions. The head of this logic program is comprised of the action name and
parameters, and the body consists of the subsets of literals from the observed delta-states that
best fits the learning examples. This process can also be implemented as a heuristic search guided
in this case by the coverage of the observed delta-states.
4. Exploitation of the learned knowledge. An action model that is built satisfying/covering a given set
of examples does not guarantee that it will present the exact set of preconditions and effects. Like
in the previous scenario, the learned models may present unnecessary preconditions. In this scenario
the learned effects may also capture invalid changes in the state caused by the partial-observability of
the pre and post states. As a consequence, planning with this kind of models may fail to address AP
problems known to be solvable.
3.2.2 Implementations
The AR MS system (Yang et al., 2007) learns PDDL planning operators from examples consisting of
tuples < s0, sn,p>; where s0is an initial state, snis a goal state and p= (a1, a2, ..., an)is a plan
corresponding to a sequence of state transitions (s0, s1, ..., sn). ARMS proceeds in two phases. In the
first phase, AR MS extracts frequent action sets from plans that share a common set of parameters. ARMS
also finds some frequent literal-action pairs with the help of the initial state and the goal state that provide
an initial guess on the actions preconditions, and effects. In the second phase, ARMS uses the frequent
action sets and literal-action pairs to define a set of weighted constraints that must hold in order to make
the plans correct. Then, AR MS solves the resulting weighted MAX-SAT problem and produces action
models from the solution of the SAT problem. This process iterates until all actions are modelled. For
a complex planning domain that involves hundreds of literals and actions, the corresponding weighted
MAX-SAT representation is likely to be too large to be solved efficiently as the number of clauses can
reach up to tens of thousands. For that reason ARMS implements a hill-climbing method that models the
10 S.JIM´
actions approximately. Consequently, the A RMS output is a model which may be inconsistent with the
(Amir and Chang, 2008) introduced an algorithm that tractably generates all the STRIPS-like models
that could have lead to a set of observations. Given a formula representing the initial belief state, a
sequence of executed actions (a1, a2, ..., an)and the corresponding observed states (s1, ..., sn)(where
partial observations of states are given), it builds a complete explanation of observations by models of
actions through a Conjunctive Normal Form (CNF) formula. By linking the possible states of fluents to
the effect propositions in the action models, the complexity of the CNF encoding can be controlled to
find exact solutions efficiently in some circumstances. The learning algorithm updates the formula of
the belief state with every action and observation in the sequence. This update makes sure that the new
formula represents all the transition relations consistent with the actions and observations. The formula
returned at the end includes all consistent models, which can then be retrieved with additional processing.
Unlike the previous approaches, the one described in (Mour˜
ao et al., 2008) deals with both missing
and noisy predicates in the observations. For each action in a given domain, they use kernel perceptrons to
learn predictions of the domain properties that change because of the action execution. LOCM (Cresswell
et al., 2009) induces action schemas without being provided with any information about initial, goal or
intermediate state descriptions for the example action sequences. LOCM receives descriptions of plans
or plan fragments, uses them to create states machines for the different domain objects and extracts the
action schemas from these state machines.
3.3 Learning Stochastic Models in Fully Observable Environments
In this scenario the observations of the state of the environment are perfect, but actions present stochastic
effects. This scenario corresponds to the probabilistic planning task. Figure 3.3 shows an example of the
unstack(B,A) action being applied in a fully observable and stochastic version of the Blocksworld. In
the example, the action produces three different alternative outcomes with probabilities of 0.5, 0.3 and 0.2
State S_1
State S_i+1
State S_i+1
State S_i+1
Figure 8 Example of application of the unstack(B,A) action in a fully observable and stochastic Blocksworld.
A Review of Machine Learning for Automated Planning 11
3.3.1 The Learning Task
1. Knowledge representation. Languages for probabilistic planning support the representation of AP
models with stochastic state-transitions. The Probabilistic Planning Domain Definition Language
(PP DDL) (Younes et al., 2005) is the current standard representation language defined for the IPC’s
probabilistic planning track. PP DDL is essentially an extension of PDD L2.1 (Fox and Long, 2003)
that supports actions with probabilistic effects. The syntax for probabilistic effects in PPD DL is
formally defined as follows:
(probabilistic p1o1p2o2. . . pkok)
where outcome oiof the action occurs with probability pi. It is required that 1pi0and
i=1 pi= 1. Figure 9 shows the PPDDL representation for the action unstack from the prob-
abilistic version of the Blocksworld defined at IPC-2004. The probabilistic effects of this action
indicate that with probability 0.5 the robot hand will successfully unstack block ?top from block
?bottom, with probability 0.3 the block ?top will fall down on the table and that with probability
0.2 the blocks will not be moved.
(:action unstack
:parameters (?top - block ?bottom - block)
:precondition (and (emptyhand) (clear ?top) (on top? ?bottom))
(probabilistic 0.5 (and (not (emptyhand)) (not (clear ?top))
(not (on ?top ?bottom))
(holding ?top) (clear ?bottom))
0.3 (and (not (on ?top ?bottom))
(on-table ?top) (clear ?bottom)))))
Figure 9 PPDDL representation for the action unstack from the Blocksworld.
2. Learning examples. Learning examples consist of tuples < si1, ai, si>where aiis the action
applied at state si1, the pre-state and siis the post-state. Given that action effects are stochastic,
pairs (pre-state,action) may have different associated post-states.
3. Learning algorithm. Because the environment is fully observable, action preconditions are captured
by lifting the pre-state of the learning examples. Nevertheless, learning stochastic action effects is
more complex because of the overlapping post-states; i.e., given an action transition defined by
the tuple (pre-state, action, post-state), several action effects could describe the transition. This
aspect of overlapping examples is also present in the induction of stochastic logic models such as
Stochastic Logic Programs (Muggleton, 1995; Cussens, 2001), Bayesian Logic Programs (Jaeger,
1997; Kersting and Raedt, 2001) or Markov Logic Networks (Richardson and Domingos, 2006).
These techniques share a two-steps strategy for implementing the model learning task:
(a) Structural Learning. At this step, ILP algorithms are used to induce first-order logic formulas
that generalize the relations from the ground facts of the learning examples.
(b) Parameter Estimation. This step estimates the probabilities associated with the induced logic
formulas. The output of the Structural Learning step is frequently a set of rules of the form
CPwhere Pis the prediction of the rule and Cis the set of conditions under which
the prediction is true. In this case, the Parameter Estimation step consists of computing the
conditional probability prob(P|C). When the induced rules are disjunctive, this probability is
directly extracted from the frequency of the learning examples but when rules overlap, estimating
this conditional probability requires more complex approaches:
i. Bayesian Estimation (BE). This approach estimates the conditional probability following
the Bayes Theorem,
prob(P|C) = prob(C|P)prob(P)
12 S.JIM´
ii. Maximum Likelihood Estimation (MLE). This approach assumes that the examples follow a
known probability distribution and iteratively tunes the distribution’s parameters to achieve
a good fit.
Accordingly, modelling an action’s stochastic effects can be defined as learning a Stochastic Logic
Program where each clause in the program represents one probabilistic effect of the action. The heads
of the clauses of the logic program consist of the action name and action parameters, and the bodies
consist of the subsets of literals from the observed delta-states that best fit the learning examples for
this effect. Figure 10 shows an example of stochastic logic program for the effects of action unstack
from the Blocksworld.
0.5:effects_unstack(TOP,BOTTOM):- object(TOP),object(BOTTOM),
0.3:effects_unstack(TOP,BOTTOM):- object(TOP),object(BOTTOM),
0.2:effects_unstack(TOP,BOTTOM):- object(TOP),object(BOTTOM).
Figure 10 Representation of the effects of action unstack from the Blocksworld as a Stochastic Logic Program.
4. Exploitation of the learned knowledge. The on-line modelling of actions with stochastic effects is
an open issue. Actions may present outcomes that result in dead-end states which complicates the
environment’s exploration. It is also not enough to detect causal flaws in the learned model. It is
necessary to detect flaws in the probabilities associated with the models. There are no mechanisms to
refine stochastic actions when states that do not fit the model are observed.
3.3.2 Implementations
In (Oates and Cohen, 1996), the authors induced propositional rules corresponding to conditional
probabilistic effects of actions. Their approach explores a given AP domain by taking random actions
and observing their execution. For each execution, they register an observation in the form of oi=<
si, ai, si+1 >, where aiis the action executed in si, and si+1 is the resulting state of executing aiin si.
Then they induce the effects of a given action by performing a general-to-specific Best-First Search in the
space of the action’s possible conditional effects. For each action, this search is guided by the frequency
of observations covered by the action’s possible conditional effects and terminates when a user-specified
number of search nodes are explored.
The TR AIL system (Benson, 1997) learns relational models of probabilistic actions. TRA IL explores
a given AP domain as follows: first it tries to solve a problem with the current action model. When
this model is too bare to solve the problem, it asks an external expert a solution plan, it observes the
plan’s execution and it learns a new model from these observations. If the current action model makes
TRAI L generate a plan for the problem, TRAIL executes that plan and observes its execution to refine
the action model. TR AIL action models are an extended version of Horn clauses so TRA IL straight off
uses standard ILP tools for learning.
The LO PE system (Garc´
ınez and Borrajo, 2000) learns planning operators with an associated
probability of success. At the beginning, the system has no knowledge. It perceives the initial state, and
selects a random action to execute in the environment. Then it loops by (1) executing an action, (2)
perceiving the resulting state and its utility, (3) learning a model from the perception and (4) planning
for further interaction with the environment (if the execution of the plan is completed, or the system has
observed a disparity between the predicted situation and the situation perceived). The LOP E’s planning
component does not explicitly receive a goal input given that LO PE creates its own goals from the highest
utility situations.
A Review of Machine Learning for Automated Planning 13
In (Pasula et al., 2007), the authors learn probabilistic action models from externally provided observa-
tions which again consist of tuples oi=< si, ai, si+1 >. The induced action models are probabilistic
STRIPS operators extended in two ways: (1) they can refer to objects not mentioned in the action
parameters, and (2) for each action, they add a noise outcome that groups the set of possible outcomes
that are the least probable for the action. The action modelling algorithm has three learning layers:
1. Learning the action structure. At this layer a greedy search process is performed in the space of action
sets to select the action preconditions and outcomes that best fit the learning examples. An evaluation
function guides the search favoring rule sets that assign high likelihood to the learning examples and
penalizing model complexity (number of preconditions plus number of outcomes).
2. Learning the action outcomes. Given an action and a set of learning examples, greedy search is used
to decide the best set of outcomes and their corresponding parameters for the action. This learning is
performed every time a new rule is constructed in the previous layer.
3. Parameter estimation. Given an action, the learning examples are used to estimate a distribution
over action outcomes. This work estimates the outcomes parameters implementing a MLE with the
conditional gradient method. The MLE is performed for each set of outcomes considered in the
previous layer.
The PELA system (Jim´
enez et al., 2008) automatically upgrades a STRIPS action model with situation-
dependent probabilities learned from plan executions. These probabilities are captured using the TILDE
tool for the automatic induction of first order decision trees (Blockeel and De Raedt, 1998).
3.4 Learning Stochastic Models in Partially Observable Environments
In this scenario actions present stochastic effects and the observations of the world state may be incomplete
and/or incorrect. Action modelling in stochastic and partially observable environments has been poorly
studied and currently, there is no general strategy for addressing this modelling task. Preliminary
work (Yoon and Kambhampati, 2007) addresses this problem using Weighted Maximum Satisfiability
techniques for finding the action model that best explains the collected observations. Specifically, in
order to learn the preconditions of every action a, this approach enumerates all the possible axioms of
the form a(x, y) : p(x, y), with all the available predicates and variable matchings. Then, it updates the
weights associated to the axioms according to a set of learning examples. Axioms corresponding to wrong
predicates or unnecessary predicates would get close to zero weight after enough learning examples, and
axioms corresponding to necessary predicates would get close to one. Likewise, for learning the effects of
every action a, this approach enumerates all the possible effects of the form a(x, y) : e(x, y)and then
updates the corresponding associated weights in a similar way.
3.5 Discussion
While learning of propositional action models has been widely studied from a theoretical approach (via
the notion of sample complexity (Strehl and Littman, 2005)), results for learning AP action models are
largely empirical. Works on model learning for AP typically evaluate their correctness by computing the
learned model’s error divergence from a given reference model. When developing planning applications
for solving problems in the real-world however, reference models are nonexistent. Some systems exploit
other kinds of AP knowledge available. The ARMS system (Yang et al., 2007) exploits a collection of
plans to evaluate the learned model’s quality. However, as previously explained, there is no guarantee that
a model that satisfies a given collection of examples succeeds in addressing a given problem.
A different approach consists of learning the model in an incremental and online manner, similar to
Reinforcement Learning. This approach does not require previous knowledge about the planning task. In
this case, we need a mechanism that guarantees convergence to a valid model. AP systems can iteratively
detect and refine defective action models after an action execution. Defective models are detected though
their differences with the observed states. When a defective action model is observed, it is refined by
collecting more experience about the action. We can solve the action inconsistency and resume the action
14 S.JIM´
learning process. As previously explained, however, the autonomous collection of AP experience is still
an open problem.
When the planning system fails to detect a defective action model, a standard planning algorithm may
not be able to find a solution plan (though it exists). Kambhampati introduced the concept of model-lite
planning (Kambhampati, 2007). This new planning paradigm advocates the development of planning
techniques that do not only search for solution plans that perfectly satisfy the current action model
but search for the most plausible solution plan according to the current action model. The core idea
is viewing a planning problem as an MPE (most plausible explanation) problem, given an incomplete
domain definition and the evidence of initial state and goals, which enables one to use publicly available
approximate solvers that scale well, e.g., MA XWALK SAT.
4 Learning Planning Search Control Knowledge
Planners’ performance can be improved by exploiting knowledge about the structure of the AP tasks
that is not explicitly encoded in the domain model. This fact was evidenced at IPC-2002 where planners
exploiting domain-specific control knowledge performed orders of magnitude faster than the state-of-the-
art planners. Hand-coding search control knowledge is complex because it implies expertise in the AP
domain and in the planner’s search algorithm. This section reviews the different systems that benefit from
ML in defining control knowledge to improve the speed and/or the quality of the planning processes.
The review is organized around the target of the learning process. We review four different approaches
in learning search control knowledge for AP (macro-actions, generalized policies, generalized heuristic
functions and hierarchical decomposition methods). The table in Figure 11 shows systems that follow
each approach. This table is not exhaustive. The systems presented are just a representative sample of
existing implementations. The following subsections explain each of the four approaches in detail.
MODEL Strengths Weaknesses
Robust to wrong learned Utility problem REFLECT (Dawson and Silklossly, 1977),
knowledge MORRIS (Korf, 1985),
Valid for different MacroFF (Botea et al., 2005a),
planners Marvin (Coles and Smith, 2007) ,
(Newton et al., 2007)
Generalized Policies
Standard relational Engineering effort to (Minton, 1988),
classification algorithms integrate with different PRIAR (Kambhampati and Hendler, 1992),
search algorithms and HAMLET (Borrajo and Veloso, 1997),
domain-independent (Khardon, 1999),
heuristics (Martin and Geffner, 2000),
DISTILL (Winner and Veloso, 2003),
OBTUSEWEDGE (Yoon et al., 2007),
CABALA (de la Rosa et al., 2007),
ROLLER (de la Rosa et al., 2008)
Generalized Heuristics
Standard relational Poor readability (Yoon et al., 2006),
regression algorithms (Xu et al., 2007)
Easy integration with
different search algorithms
and heuristics
Decomposition Methods
Very expressive No completely CAMEL (Ilghami et al., 2002),
automatic learning HDL (Ilghami et al., 2006),
HTNMAKER (Hogg et al., 2008)
Figure 11 Overview of AP systems that benefit from ML for the extraction of domain-specific search control.
4.1 Macro-actions
Macro-actions were among the first attempts to speed-up planning processes. They are extra actions
resulting from assembling actions that are frequently used together. Macro-actions reduce the depth of
the planning algorithm’s search tree. Their benefit decreases with the number of new macro-actions as
they enlarge the search tree’s branching factor causing the utility problem. Addressing this trade-off is
A Review of Machine Learning for Automated Planning 15
the key issue at stake when attempting to make macro-actions work. Figure 12 shows an example of
the macro-action unstack-putdown for the Blocksworld domain. This macro-action is the result of
assembling actions unstack and putdown.
;;; Primitive actions
(:action unstack
:parameters (?x - block ?y - block)
:precondition (and (on ?x ?y) (clear ?x) (handempty))
:effect (and (holding ?x)
(clear ?y)
(not (clear ?x))
(not (handempty))
(not (on ?x ?y))))
(:action putdown
:parameters (?x - block)
:precondition (holding ?x)
:effect (and (not (holding ?x))
(clear ?x)
(ontable ?x)))
;;; Macro-action
(:action unstack-putdown
:parameters (?x - block ?y - block)
:precondition (and (on ?x ?y) (clear ?x) (handempty))
:effect (and (clear ?y)
(not (on ?x ?y))
(ontable ?x)))
Figure 12 Macro-action unstack-putdown for the Blocksworld domain.
4.1.1 Learning Macro-Actions
1. Knowledge representation. Macro-actions are represented as new actions of the action model so
they follow the predicate logic representation of AP actions. Actions aiand ajare assembled into a
macro-action ai,j in the following way:
The parameters par(ai,j )(par(ai)Spar(aj)).
The preconditions pre(ai,j )=(pre(ai)Spre(aj)) \add(ai).
The positive effects add(ai,j)=(add(ai)Sadd(aj)) \del(aj).
The negative effects del(ai,j )=(del(ai)Sdel(aj)) \add(aj).
2. Learning examples. Learning examples are solution plans pthat consist of sequences of instantiated
actions p= (a1, a2, ..., an)that correspond to the sequences of state transitions (s0, s1, ..., sn)such
that s0is the initial state of the problem, siresults from executing action aiat state si1and snis a
goal state, i.e., Gsn, where Gare the problem goals.
3. Learning algorithm. Algorithms for learning macro-actions extract action sub-sequences from the
solution plans and count their occurrences to capture the most useful ones. Typically, the process to
extract action sub-sequences defines two parameters:
l, the length of the macro-actions. The minimum value for lis l= 2 because a macro-action
should have at least two actions. The maximum value for lis the length of the longest solution
plan in the learning examples. In practice this value has to be smaller to be useful.
k, the number of actions that can be skipped from a solution plan. This parameter allows the
learning system to extract macro-actions that ignore, at most, kirrelevant intermediate actions
from a solution plan. In practice, small values for kreduce the number of macro-actions but k
values that are too small may skip useful macros-actions occurrences.
Regarding these parameters, the complexity of extracting macro-actions of length lfrom a plan with
nactions is l+k
ln. The first factor is the cost to enumerate macro-actions of length lwithin a l+k
16 S.JIM´
sized window. The second factor is the cost of sliding the window along the solution plan. A similar
approach can be used to extract macro-actions from partially-ordered plans (Botea et al., 2005b).
4. Exploitation of the learned knowledge. Macro-actions are usable straight away by any off-the-
shelf planner because they can be included in the action model as standard actions. The inclusion
of macro-actions can damage the time and quality performance of the original action model. This
happens when macro-actions produce a reduction in the depth of the search that does not payoff
the increase in the branching factor (the utility problem). Current systems for learning macro-actions
evaluate this problem experimentally. For example, they define a set of AP problems similar to the
target ones and if a learned macro-action improves the planner’s performance in these problems, then
it is considered a good candidate to populate the action model. This approach for dealing with the
utility problem was first introduced in systems that learn control rules (Minton, 1988).
4.1.2 The Implementations
Since the beginning of AP the use of macro-actions has been widely explored. The first macro-action
learning system was STRIPS (Fikes et al., 1972). It used previous solution plans as macro-actions
to solve subsequent problems as well as to monitor the execution of plans in the real world. Later,
MORR IS (Korf, 1985) extended this approach by adding filtering heuristics to prune the generated set of
macro-actions. This approach distinguished two types of macro-actions: S-macros that occur frequently
during search and T-macros, that occur less often, but model some weakness in the heuristic. The
REFL ECT system (Dawson and Silklossly, 1977) took the alternative approach of generating macro-
actions based on domain pre-processing. All sound combinations of actions were considered to be macro-
actions and filtered through basic pruning rules. Due to the small size of the domains with which the
planner was reasoning, the remaining number of macro-actions was small enough to use in planning.
Macro-actions systems traditionally use an off-line approach to generate and filter macro-actions before
using them but a few systems have tried ML to filter macro actions during the search. In (McCluskey,
1987) the author used chunks and in (Garc´
an et al., 2006) they used control rules to decide when to
apply the macro-actions.
Recent works successfully integrate macros with state-of-the-art heuristic search planners. These works
include the IPC-2004 competitor MA CROFF (Botea et al., 2005a, 2007). Here partially ordered macro-
actions are extracted by identifying statically connected abstract components and then, an off-line filtering
technique is used to prune the list of macro-actions. Marvin (Coles and Smith, 2007), also a participant
at IPC-2004, used action-sequence memorization techniques to generate macro-actions on-line that allow
the planner to escape from plateaus without any exploration. Wizard (Newton et al., 2007) used a genetic
algorithm to generate and filter a collection of macro-actions independent from the baseline planner.
Algorithms for n-grams analysis have also been recently applied to learning macro-actions for the heuristic
planner FF (Muise et al., 2009).
4.2 Generalized Policies
A generalized policy is a mapping of planning contexts (sometimes also called meta-states) into the
preferred actions to apply in the context. A planning context typically consists of the current state as well
as with the set of goals. An accurate generalized policy is able to solve any problem from a given domain
without any search having to be done, by repeatedly applying the preferred action for each planning
context. Figure 13 shows an example of a generalized policy in the form of a decision list (ordered set
of decision rules) for the Blocksworld domain. This generalized policy stacks block TOP on block
BOTTOM when TOP should be on BOTTOM,BOTTOM is clear and BOTTOM is already well placed 3. In all
other cases the policy unstacks all blocks and puts them down on the table so they can be placed in
their final position.
3The concept of well-placed should be given in the domain theory. A block is considered well-placed when it is on
the right block (or table) with respect to the goal, and all blocks beneath it are well-placed too.
A Review of Machine Learning for Automated Planning 17
stack(Top,Bottom) :- state_holding(TOP), goal_on(TOP,BOTTOM), state_clear(BOTTOM),
putdown(BLOCK) :- state_holding(BLOCK).
unstack(TOP,BOTTOM) :- state_clear(TOP), state_armempty().
pickup(BLOCK) :- state_clear(BLOCK), state_armempty().
Figure 13 A generalized policy in the form of a decision list for the Blocksworld domain.
4.2.1 Learning Generalized Policies
Knowledge representation. A key issue when representing planning contexts is selecting the feature
space. The feature space is the set of predicates considered for the learning process. This set must be
general enough to capture the domain’s relevant knowledge and specific enough to make the learning
tractable. In AP the feature space has usually consisted of predicates for describing the current state
and the goals of the planning task. The PRODIGY system enriched the feature space with extra
predicates, called metapredicates (Minton, 1988). Metapredicates capture useful knowledge about
the planning context, like the current applicable operators or the goals that still need to be achieved.
Recent works on learning generalized policies have extended the definition of metapredicates to
include concepts from heuristic planning, like metapredicates for capturing the actions of the relaxed
plan in the current state (Yoon et al., 2008) or for capturing the set of helpful actions in the current
state (de la Rosa et al., 2008).
There have been two main approaches for representing generalized policies. A generalized policy can
be represented as a set of rules that capture the preferred action to apply in a given planning context.
In this case, the generalized policy is formally defined as a tuple π=< L, R >, where
Lthe set of literals used to describe the different planning contexts. It defines the feature space
used during learning.
Ris a set of rules, where the rule’s Head is the action to apply and the Body is the set of
predicates that describes the planning context under which the action should be applied.
A generalized policy can also be represented as a set of decision instances with a distance metric
to retrieve similar instances. This is the representation approach that CBR (Case-Based Reasoning)
planners follow. Even if this representation is able to capture more specific domain regularities, it has
the serious drawback of requiring an appropriate similarity metric. Utility is also a problem when
handling large collections of decision instances, because, as a collection gets larger, the time it takes
to search for similar instances also increases. This second representation for generalized policies is
formally defined as a tuple π=< L, I, D > where:
Lthe set of literals used to describe the planning contexts. It defines the feature space.
Iis a set of tuples, i=< ci, ai>, where ciis an instantiated planning context and aiis the
instantiated action applied in ci.
Dis a distance metric that computes the distance between two different planning contexts. Given
a new planning context c, the policy decides the action aito execute in cby computing the closest
tuple in Iand returning its associated action ai, as shown in,
π(c) = argaimin
(<ci,ai>I)D(c, ci)
Both approaches usually represent planning contexts in predicate logic because it is a natural
encoding for the AP tasks. Predicate logic provides mechanisms to define extra predicates that enrich
planning contexts. As an example of these extra predicates, Figure 14 shows the definition of the
wellplaced(Block) predicate for the Blocksworld (Khardon, 1999). This concept is not in the
Blocksworld domain’s original coding, but it is useful to define a compact generalized policy as is
shown in the example of Figure 13.
18 S.JIM´
ingoal(Block) :- goal_ontable(Block).
ingoal(Top) :- goal_on(Top,Bottom).
wellplaced(Block) :- state_ontable(Block), goal_ontable(Block).
wellplaced(Block) :- state_on(Block,Bottom), not(ingoal(Bottom)).
wellplaced(Block) :- state_on(Block,Bottom),goal_on(Block,Bottom),wellplaced(Bottom).
Figure 14 Representation of the wellplaced(Block) concept for the Blocksworld domain in predicate logic.
Learning these kinds of predicates in predicate logic is still an open issue so they have to be hand-
coded. Planning systems have used other representation languages that succeed in learning these
concepts. Languages for describing object classes have been shown to provide a useful bias for
learning these concepts. These languages include concept language (Martin and Geffner, 2000) and
taxonomic Syntax (Mcallester and Givan, 1989) which provide operators to define recursive concepts
over predicates, like the operator. In the Blocksworld domain, for example, the useful concept of
well-placed block can be defined in a very compact way using these languages (Figure 15).
(goal ontable =state ontable) : wellplaced
(goal on.(goal on =state on)) : wellplaced
Figure 15 Well-placed concept for the Blocksworld domain expressed in concept language.
Expressing planning knowledge in temporal logic has also been shown to be useful. This is the case
of TLPL AN (Bacchus and Kabanza, 2000). Unfortunately, learning planning knowledge in temporal
logic has not been resolved.
Learning examples. Learning examples are extracted from solutions to training problems. The
learning examples consist of tuples < ci, ai+1 >where ciis a planning context and ai+1 is the action
applied at context ci.
Learning algorithm. When the policy consists of a set of rules, the learning task is very much related
to the ILP task. Learning a generalized policy can be defined as the induction of a logic program for
every action in the domain. This logic program captures when to apply the action. The head of the
rules in this logic program consists of the action name and parameters, and the body is the subset
of literals from the planning context that best covers the learning examples. This learning task can
be implemented as a heuristic search guided by the coverage of learning examples. When the policy
consists of a set of relevant instances, the learning task stores and manages the set of instances. In
this case, learning a large number of instances can actually be counterproductive because it is hard to
store and manage them and because of the difficulty involved in determining which instance to use to
solve a particular problem. A solution to this problem is to post-process the stored instances to only
manage the most relevant ones. Examples are REPLIC A (Garc´
an et al., ress) which extracts
a set of prototypes from the instances or DI STILL (Winner and Veloso, 2003) which builds a highly
compressed instances library by generalizing and merging solution plans.
Exploitation of the learned knowledge. A generalized policy can be used to directly select which
action to apply for a given planning context. When a learned generalized policy is imperfect, however,
its application may fail to solve problems like in the case of the early systems in this category.
These systems learned control-rules to guide the exploration of planning search trees. A control-
rule is an IF-THEN rule that proposes node selection, pruning or ordering during tree exploration.
A set of control rules can be viewed as a partial generalized policy since they do not provide action
recommendation for all possible planning contexts. When a given control-rule recommendation was
imperfect, it often prevented the planner from finding a solution plan. To exploit generalized policies
A Review of Machine Learning for Automated Planning 19
more robustly, recent works have exploited the learned policy as an evaluation function within search
algorithms like Beam-Search or Limited Discrepancy Search. These search algorithms allow planning
to combine the guidance provided by the learned knowledge with other sources of advice like domain-
independent heuristics.
4.2.2 Implementations
There are a group of systems that inductively learn control rules. Among them Inductive Learning
Programming (ILP) is the most popular learning technique. The GRA SS HOP PE R system (Leckie and
Zukerman, 1991) used FOIL (Quinlan and Cameron-Jones, 1995) to learn control rules that guide the
PRODIGY planner. There are also analytical systems: PRODIGY/EB L module (Minton, 1988) learned
search control rules for the PRODIGY planner from a few examples of right and wrong decisions.
STATIC (Etzioni, 1993) obtains control rules without solving any problems. It simply uses Explanation
Based Learning (EBL) to analyze the relations between actions preconditions and effects. With the aim
of overcoming the limitations of purely inductive and purely analytical approaches some researchers tried
to combine them: the pioneering systems based on this principle are LEX-2 (Mitchell et al., 1982) and
META- LEX (Keller, 1987). AXA-EBL (Cohen, 1990) combined EBL and induction. It first learns control
rules with EBG and then refines them with learning examples. DO LP HIN (Zelle and Mooney, 1993; Estlin
and Mooney, 1996) was an extension of AXA-EBL which used FOIL as the inductive learning module.
The HAMLET (Borrajo and Veloso, 1997) system combines deduction and induction incrementally. First it
learns control rules with EBL that usually are too specific or general and then uses induction to generalize
and specialize the rules. EVOCK (Aler et al., 2002) uses Genetic Programming to evolve the rules learned
by HAMLET and yield more effective search control.
The problem of learning generalized policies was first studied by Roni Khardon. Khardon’s
L2ACT (Khardon, 1999) induced generalized policies for both the Blocksworld and the Logistics domain
by extending the decision list learning algorithm (Rivest, 1987) to the relational setting. This first approach
presented two important weaknesses: (1) it relied on human-defined background knowledge that expressed
key features of the domain, e.g. the predicates above(block1,block2) or in place(block) for
the Blocksworld, and (2) the learned policies did not generalize well when the size of the problems
increases. Martin and Geffner solved these limitations in the Blocksworld domain by changing the
representation language of the generalized policies from predicate logic to concept language to learn
recursive concepts (Martin and Geffner, 2000).
Recently the scope of generalized policy learning has increased over a variety of domains making
this approach competitive with state-of-the-art planners. This achievement is due to two new ideas: (1)
the policy representation language is enriched with extra predicates that capture more effective domain-
specific knowledge; and (2) the learned policies are applied not greedily but within the framework of
heuristic search algorithms. A prominent example is the OBTUSEWE DG E system (Yoon et al., 2007),
the best learner of the IPC-2008 learning track. This system enriches the knowledge of the current state
with the relaxed planning graph and uses the learned policies to generate look-ahead states within a Best-
First Search. ROLL ER (de la Rosa et al., 2008) defined the problem of learning generalized policies as
a two-step standard classification process. In the first step a classifier captures the preferred operator to
be applied in the different planning contexts. In the second one, another classifier captures the preferred
bindings for each operator in a given domain’s different planning contexts. These contexts are defined by
the set of helpful actions extracted from a given state’s relaxed planning graph, the goals still unachieved,
and the the planning task’s static predicates.
Systems also exist that represent generalized policies through collections of planning instances, as
CBR systems for AP. Instance-based AP systems are not usually competitive in a variety of domains
because they present the drawback of having to define an appropriate similarity metric that works well in
different kinds of domains. The PRIAR system (Kambhampati and Hendler, 1992) proposed integrating
the modification of plans with generative planning. PRO DI GY/A NA LO GY (Veloso and Carbonell, 1993)
introduced the application of derivational analogy to planning. It stored planning traces to avoid failure
paths in future problem solving episodes. To retrieve similar planning traces, PRO DIG Y/A NA LO GY
20 S.JIM´
indexed them using the minimum preconditions to reach a set of goals. The cased-based planning system
PARIS (Bergmann and Wilke, 1996) proposed the introduction of abstraction techniques to store the cases
organized in a hierarchical memory. This technique improves the flexibility of the cases adaptation, thus
increasing the coverage of a single case. DIST ILL (Winner and Veloso, 2003) incorporates example
plans into a structure called dsPlanner. DISTILL converts the plan into a parameterized if-statement
and searches through each of the if-statements already stored in the dsPlanner to merge them. If the
learned dsPlanner is accurate it can be used directly to solve any problem from the domain without
search. CA BALA (de la Rosa et al., 2007) uses object centered solutions plans, called typed sequences,
for node ordering during the plan search in heuristic planning. Another instance-based learner, REPL IC A,
implemented a Nearest Prototype Learning (Garc´
an et al., ress) using a relational distance inspired
by the metrics used in relational data-mining. OAKPlan (Serina, 2010) uses a compact graph structure
to encode planning problems. This structure provides a detailed description of the planning problem’s
topology and allows the learner to define an effective retrieval procedure based on kernel functions.
In (de la Rosa et al., 2009) a planning system was recently described that was able to benefit from
generalized policies represented in any of the two approaches (rules or instances). This system queries
the policies for fixing the flaws of relaxed plans, and then it uses the resulting relaxed plans as grounded
macro-actions in a Best-First Search.
4.3 Generalized Heuristic Functions
Heuristic functions are used in AP to focus the search of a solution plan on the search nodes that seem the
most promising. Heuristic functions for AP compute an estimate of the distance from a given search node
to a node in which goals are satisfied. Domain-independent heuristic functions for AP can be directly
derived from the cost of the solutions to a relaxed task. The most common relaxation of the AP task
consists of ignoring the delete effects of actions. Most of today’s heuristic planners rely on this idea
to implement their heuristics. Since this approach is domain-independent, it fails to capture planning
domains singularities. This section describes how to obtain AP heuristics that capture domain-specific
knowledge with ML. It focuses on heuristics for the most popular search approach in AP, the forward
state-space search.
4.3.1 Learning Generalized Heuristic Functions
Knowledge representation. They are functions H(s;A;G)of a state s, the action model Aand a
goals set G, that estimate the cost of achieving the goals Gstarting from sand using actions from
A.Predicate logic is a natural encoding for AP heuristic functions because they express knowledge
about the current state, goals and actions of the AP task. Representation languages, like taxonomic
Syntax, that focused on the properties of objects have also been used. Figure 16 shows an example
of a simple domain-specific heuristic for the Blocksworld. This heuristic function estimates a cost
value of 0, 1, 4 or 6 steps to achieve the goal on(A,B) with regard to different configurations of the
current state.
+ yes: [0]
+--no: on(C,A) AND on(D,B)?
+--yes: [6]
+--no: on(C,A) OR on(D,B)?
+--yes: [4]
+--no: [1]
Figure 16 Domain-specific heuristic for the Blocksworld.
Learning examples. Learning examples consist of tuples < si, ci, gi>where siis the current state
and ciis the actual cost of achieving the goals gifrom the state si.
A Review of Machine Learning for Automated Planning 21
Learning algorithm. The aim of the learning algorithm is generalizing the cost values captured
in the learning examples. Since the target concept to be generalized from the learning examples is
numeric, this learning task corresponds to a regression task. When learning examples are represented
in predicate logic, standard relational regression algorithms can be used, such as learning relational
regression trees (Blockeel et al., 1998).
Exploitation of the learned knowledge. The main advantage of this approach is that the learned
knowledge can be directly combined with other standard sources of guidance for AP such as domain-
independent heuristics. The learned knowledge is difficult for humans to understand.
4.3.2 Implementations
In (Yoon et al., 2006), the authors build a generalized heuristic function through linear regression.
They learn domain-specific corrections ∆(s;A;G) = Piwifi(s;A;G)to the relaxed plan heuristic
RP H (s;A;G)introduced by the FF planner (Hoffmann and Nebel, 2001b). These corrections are
expressed as a weighted linear combination of features, where wiare the weights and firepresent the
different features of the planning context. The regression examples consist of observations of the true
distance to the goals from different states obtained with the FF planner. The resulting heuristic function,
H(s;A;G) = RP H (s;A;G) + ∆(s;A;G)
provides more accurate estimates that capture domain-specific regularities.
The previous approach ignores the actual performance of the heuristic when used in a search algorithm.
It may attempt to learn corrections for the heuristic even if it provides good guidance during search. In (Xu
et al., 2007), they present an alternative approach that only considers learning when the heuristic misleads
in a given search strategy. The heuristic learned by this approach,
H(s;A;G) = X
only attempts to discriminate between good and bad states well enough to find the goals within a Beam-
Search procedure, rather than attempting to model the distance to the goals precisely. This approach
implements the following weight learning strategy. For each state sjin the solution to a learning problem,
if sjis not contained in the search’s beamj, there is a search error. In this case, weights are updated like
the Perceptrons weights, so that state sjwill be preferred by the heuristic, remaining in the beamjin
future search episodes.
4.4 Hierarchical Decomposition Methods
An extended strategy for coping with problems complexity is to decompose problems into simpler
sub-problems. When an effective decomposition is found, the sum of the costs of addressing the sub-
problems is smaller than the cost of directly addressing the original problem. Hierarchical Task Networks
(HTN) is one of the best studied approaches for modelling AP task decomposition. The HTN approach
combines hierarchical domain-specific representations of the planning task and domain-independent
search strategies for problem solving.
The input to a HTN planner includes an action model that encodes a set of primitive actions similar to
the ST RIPS actions used in classical planning and a task model. The task model defines a set of methods
which describes how tasks should be decomposed into subtasks in a particular domain. The job of the HTN
planner consists of exploiting the task model to decompose a given planning task into simpler subtasks
until a sequence of primitive actions is generated. Figure 17 shows the primitive action unstack and
the method for the task move-block from a HTN description of the Blocksworld domain. This method
moves a block to its final position by defining two subtasks: moving-x-from-y-to-z that unstacks
block xfrom y, in order to stack xon z, and moving-x-from-table-to-z that picks-up block x
from the table and stacks it on z. In both subtasks, the method checks that the block xcan be directly
moved to its final position (primitive actions check,check2 and check3) and makes sure that xwill not be
moved in the future.
22 S.JIM´
(:operator (!unstack ?top ?bottom)
((clear ?top) (on ?top ?bottom))
((holding ?top) (clear ?bottom)))
(:method (move-block ?x ?z) ;head
(:first (on ?x ?y))
; Decomposition
((!unstack ?x ?y) (!stack ?x ?z)
(!assert ((dont-move ?x)))
(!remove ((stack-on-block ?x ?z)))
(check ?x) (check2 ?y) (check3 ?z))
; Decomposition
((!pickup ?x) (!stack ?x ?z)
(!assert ((dont-move ?x)))
(!remove ((stack-on-block ?x ?z)))
(check ?x)))
Figure 17 Examples of operator and method for HTN planning in the Blocksworld domain.
Current HTN planners, such as SHOP2 (Nau et al., 2003), can outperform state-of-the-art domain-
independent planners and provide a natural modelling framework for many real-world applications like
fire extinction (Castillo et al., 2006), evacuation planning (Mu˜
noz-Avila et al., 1999) or game playing (Nau
et al., 1998). Defining effective decomposition methods is still complex because these methods reflect deep
knowledge of the domain. The operator and method shown in Figure 17 belong to a HTN representation
of the block-stacking algorithm (Erol et al., 1992). Below we review some learning approaches to
automatically define decomposition methods within the HTN planning paradigm.
4.4.1 Learning Hierarchical Decomposition Methods
1. Knowledge representation. A HTN method is a procedure that describes how to decompose a non-
primitive task into primitive tasks or simpler non-primitive tasks. A method mis formally defined
by a triple m=<head, preconds, subtasks>, where head represents the name and parameters of the
task to decompose, preconds is a logical formula denoting the method preconditions, and subtasks is
a partially ordered sequence of subtasks. A method mis applicable to a state sand task tif head(m)
matches tand the preconds(m)are satisfied in s. The result of applying a method mto a state sand
task tis the sequence of subtasks subtasks(m).
2. Learning examples. Learning examples consist of a set of planning problems, that is a set of pairs
< s0,G>where s0is an initial state and Gis a set of goals, together with the corresponding solution
plans p= (a1, a2,...,an). These solution plans can be human provided or generated by a classical
3. Learning algorithm. The algorithms reviewed in Section 3 to learn the action model preconditions
can also be applied to learn the HTN method preconditions. A reduced version of the HTN method
decompositions can be induced by using hierarchical layers of macros. Decompositions induced
following this approach do not exploit the HTNs full expressive power that includes definitions
of alternative decompositions, recursive concepts or loops. At this point, there is no completely
automatic algorithm for learning subtask decompositions that benefits from HTNs full expressiveness.
Some algorithms need to start from decomposition methods partially specified by humans, others need
hierarchical plans (which implies previously specified methods) and some others need certain tasks
called annotated tasks specified that define the equivalence of the HTN tasks and the set of goals G
in a classical version of the AP problem.
A Review of Machine Learning for Automated Planning 23
4. Exploitation of the learned knowledge. The generality level of the learned knowledge is the key
issue to make a given HTN description effective. Learning decomposition methods that are too
general may produce infinite recursion when problem solving, decreasing the advantages of HTN
planning over classical planning. On the other hand, learning too specific HTN methods is less likely
to be effective in decomposing new problems. Current methods to address this trade-off are based on
measuring the performance of the learned model against a set of test problems.
4.4.2 Implementations
CAMEL (Ilghami et al., 2002) learns the preconditions of HTN decomposition methods from observations
of plan traces, using the version space algorithm (Mitchell, 1997). CAMEL is designed for domains where
the planner receives multiple methods per task, but not their preconditions. The structure of the hierarchy
here is known in advance and the learning task tries to identify under which conditions the different
hierarchies are applicable. This approach requires many plan traces to completely converge (completely
determine the preconditions of all methods). CAMEL++ (Ilghami et al., 2005) enables the planner to start
planning before the method preconditions are fully learned. By doing so, the planner can start solving
planning problems with a smaller number of training examples. Both CAMEL and CAMEL++ require
each input plan trace needs to contain extra information so that the learner could acquire the models. At
each decomposition point in a plan trace, the learner needs to have all the applicable method instances,
not just the one that was actually used.
The HTN Domain Learning (HD L) algorithm (Ilghami et al., 2006) starts with no prior information
about the methods. It examines hierarchical plan traces produced by an expert problem-solver. For each
decomposition point in the traces, HD L checks to see if the responsible method is already present. If not,
HDL creates a new method and initializes a new version space to capture its preconditions. The method
that was actually used to decompose the corresponding task serves as a positive example and the methods
that matched that task but whose preconditions failed, serve as negative examples for the corresponding
version space.
HTN-Maker (Hogg et al., 2008) generates an HTN domain model from a ST RIPS domain model, a
collection of plans pgenerated by a ST RIPS planner, and a collection of annotated tasks. An annotated
task is a triple (n, Pre, Effects) where nis a task, Pre is a set of atoms known as the preconditions, and
Eff is a set of atoms known as the effects. HTN-Maker first generates the list of states S= (s0,...,sn)
by applying the actions in the plan pstarting from the initial state s0. Then, it traverses these states, and
if there is an annotated task whose effects match the state si+nand whose preconditions match the state
si, it regresses the effects of the annotated task through (1) previously learned methods or (2) a primitive
task if there are no previously learned methods.
4.5 Learning Planning Search Control in Domains with Uncertainty
When planning in domains with uncertainty, the combinatorial explosion problem becomes even more
acute. Partial observability of the environment and non-deterministic outcomes of actions multiply the
number of states that can be reached, and increase the complexity of the search processes. Search control
knowledge can help reduce this complexity but as mentioned before, hand-coding this knowledge is
difficult and requires expertise on the domain and the planner. There is a review of techniques for Planning
Under Uncertainty (PUU) that improve their own scalability through ML. These techniques preserve the
representation formalisms and learning algorithms reviewed in the section but adapt the mechanisms to
generate the learning examples and apply the learned knowledge to the PUU framework:
Macro-actions. In (Theocharous and Kaelbling, 2003), the authors solve large Partially Observable
Markov Decision Processes (POMDP) by using a dynamically-created finite-grid approximation of
the belief space and then applying dynamic programming to compute a value function at the grid
points. This work exploits macro-actions to visit a significantly smaller part of the belief space of the
POMDP than with simple primitive actions. Macro-actions also make the convergence of the dynamic
programming faster because the agent can look further into the future.
24 S.JIM´
Generalized policies. Given a set of small probabilistic planning problems, the work presented
in (Yoon et al., 2002) uses the probabilistic planner P GRA PH PLA N to synthesize solution plans
p= (a1, a2,...,an)corresponding to a sequence of state transitions (s0, s1, ..., sn)such that si
results from executing action aiin the state si1and snis a goal state. The execution of plans p
are simulated one action at a time. If the state resulting from simulating aiis different from si, then
a new plan is computed to attain the goals in the new state. This process is repeated until a goal
state is reached or a limit of simulations is exceeded. A generalized policy is then learned from the
pairs (state,action) traversed in the simulations. In many domains, solutions to small problems do not
capture the singularities of larger problems. As an alternative, in (Fern et al., 2006) the authors learn
the generalized policy from pairs (state, action) generated following the Approximate Policy Iteration
algorithm (Bertsekas and Tsitsiklis, 1996).
Generalized evaluation functions. Given a set of small probabilistic planning problems, the approach
presented in (Gretton and Thi´
ebaux, 2004) computes the optimal policies for the problems with
a MDP solver. Then, it uses first-order regression to generalize a first-order value function over
the values of optimal policies. In (Sanner and Boutilier, 2005, 2006) the authors solve First Order
MDPs (FOMDPs) by representing the value function linearly with respect to a set of first-order basis
functions and computing suitable weights by casting the corresponding optimization as a first-order
linear program.
Decomposition methods. In (Hogg et al., 2009), the HTN-MAKE R algorithm has been adapted to
non-deterministic planning domains, where actions may have multiple possible outcomes.
4.6 Discussion
IPC-2008 created a specific track for AP systems that benefit from learning search control. This
competition was the first forum to compare the performance of different planning and learning systems in
a common test-bench. Most of the participants based their proposals on heuristic planning and presented
relevant advances in the field like extensions to the planning context with information from the relaxed
planning graph or techniques to combine learned knowledge with other sources of guidance like domain-
independent heuristics. One of the most important conclusions drawn from the competition was that none
of the participant systems is successful in capturing useful knowledge in all the proposed domains. In
some domains, the representation language’s bias prevented the systems from learning this knowledge, in
others the cause was the learning examples’ poor quality or the learning algorithms’ limitations. It became
clear that the best approach for learning useful knowledge in a given domain need to be ascertain. This
task involves better understanding:
1. What the best representation for a given domain is. IPC-2008 showed that action policies are inaccu-
rate in domains like Sokoban and N-puzzle because they look over knowledge regarding the trajectory
on the way to the goals. Numeric functions, like the Manhattan distance that provides a lower-bound
for the solution’s length, produce useful search control for these domains. Further studies are needed
to establish theoretical considerations about the expressive power and/or limitations of the various
forms of learned knowledge. These studies should also cover other kinds of planning paradigms able
to exploit more expressive representations than off-the-shelf planners like planners using temporal
logic (Bacchus and Kabanza, 2000), or AP systems able to reason about loops (Srivastava et al.,
2008; Winner and Veloso, 2003; Shavlik, 1989; Larkin et al., 1986).
2. What the best learning examples for a given domain are. In most of the systems reviewed, the
learning examples are extracted from experience collected solving problems from a training set. These
examples can be collected in different ways to suit the learning task’s singularities:
(a) Quality. In order to learn search control knowledge that provides accurate guidance towards the
goals, learning examples should correspond to good quality solutions. Since optimal algorithms
for AP are expensive in terms of time, a more practical approach for generating good solutions
consists of using Branch and Bound strategies (de la Rosa et al., 2009).
A Review of Machine Learning for Automated Planning 25
(b) Diversity. Some domains, like logistics, present many solutions for a given planning problem that
correspond to swaps in the order of certain actions. These solutions express different decisions
for the same planning context, which makes learning more complex. In order to avoid this effect,
learning can be completed from partially ordered plans (Veloso et al., 1990). Partially-ordered
plans contain the same set of actions as the totally-ordered plans, but partially-ordered plans only
include the necessary constraints on the action-ordering.
(c) Eagerness. Learning examples can correspond to every planning decisions (eager approach) or
just to decisions where the planner failed to make the desired choice (lazy approach), i.e., the
planner choice was not part of the solution or was not optimal. A lazy approach is planner-
dependent, but it generates more compact search control knowledge.
3. What the best learning algorithm for a given domain is. Different learning algorithms present different
biases that affect the learning process like where decision lists learning algorithms specify the
maximum number of literals per rule. Different learning algorithms can also present different memory
and learning time performance and, depending on the algorithm, the resulting model can present a
different readability and applicability. Lazy learning algorithms present low learning times but large
times when exploiting that learned knowledge. Recently, works on meta-learning (Brazdil et al., 2009)
try to compute estimates that characterize these features of the learning algorithms but they normally
focus on propositional versions of the ML algorithms.
4. When to exploit the learned knowledge. Most of the participants in the learning track of IPC-2008
suffered from performance problems caused by the use of imperfect learned knowledge. In practice,
there is no sign of a given piece of search control knowledge’s quality and the only possible evaluation
is to compare the performance of the planning system with and without the learned knowledge.
This evaluation by testing strategy is the underlying motivation of AP systems based on portfolio
of planners (Howe et al., 1999; Vrakas et al., 2005). These systems evaluate different planning
configurations in a set of training problems to obtain the planning setup that allows them to address
the target problems more robustly. This approach was also followed by the winner of the IPC-2008
learning track (Gerevini et al., 2009a).
5 Reinforcement Learning
RL agents interact with the environment to collect experience that– with the appropriate algorithms–
is processed to generate an optimal policy (Kaelbling et al., 1996; Sutton and Barto, 1998). Building
a RL agent involves similar decisions to the ones in learning for AP: the representation (how to
encode the agent’s environment and actions); the learning examples (the agent’s strategy for collecting
experience); the learning algorithm (what kind of algorithm performs the best for the specified task); and
the exploitation of the learned knowledge (how can the agent benefit from the obtained knowledge). Unlike
most of the learning for AP approaches, RL present tightly integrated solutions for knowledge acquisition
and knowledge exploitation. A key issue in these integrated solutions is determining when to try new
actions and when to use known ones. RL studies this issue as the exploration-exploitation dilemma, where
exploration is defined as trying new actions and exploitation is defined as applying actions that succeeded
in the past. In general terms, good answers to the exploration-exploitation dilemma consider the number
of trials allowed; the larger the number of trials is, the worse is to prematurely converge to known actions
because they may be suboptimal. For surveys on efficient exploration-exploitation strategies see (Wiering,
1999; Reynolds, 2002).
5.1 Model-Based and Model-Free RL
RL was primarily concerned with obtaining optimal policies when the model of the environment is
unknown but actually, there are two ways to proceed. Model-Based RL requires a environment’s model
consisting of a transition and a reward model. Model-Based RL typically relies on standard dynamic
programming algorithms (Bellman and Kalaba, 1965; Bertsekas, 1995) to find a value/heuristic function
that provides optimal policies. Learning in Model-Based RL is understood as in Real-Time Heuristic
26 S.JIM´
search (Korf, 1990; Bulitko and Lee, 2006; Hern´
andez and Meseguer, 2007), i.e., as local updates of the
value/heuristic function with information obtained from simulation.
Model-Free RL does not require a model of the environment. Below, there are two different approaches
for the Model-Free RL task:
Combining learning action models with Model-Based RL. This approach learns a transition model of
the environment and applies standard dynamic programming algorithms to find a good policy.
Pure Model-free RL. In domains with large uncertainty learning to achieve goals is easier than
learning a model of the environment. Pure Model-free RL algorithms do not model the decision-
making as a function of the state, like value/heuristic functions, but with a function of pairs
< state, action > called action-value functions. The Q-function, which provides a measure of the
expected reward for taking action aat state s, is an example of an action-value function.Q-
learning (Watkins, 1989), a well-known Pure Model-free RL algorithm, updates the Q-function with
every observed tuple < s, a, s0, r > (s0represents the new state and rrepresents the obtained reward).
Q-learning completes the update of the Q-function using the following formula,
Qt+1(s, a) := Qt(s, a) + α(r(s, a) + γmax
a0Q(s0, a0)Qt(s, a))
where αis the learning rate that determines to what extent the newly acquired information overrides
the old information. When α= 0 the agent does not learn anything, while when α= 1 the agent only
considers the most recent information. γis the discount factor which determines the importance of
future rewards. A factor of γ= 0 makes the agent greedy (the agent only considers current rewards),
while a factor approaching 1 makes the agent strive for a long-term high reward. Pure model-free RL
also includes Monte Carlo methods (Barto and Duff, 1994). Most of pure model-free RL methods
guarantee to find optimal policies and use very little computation time per observation. However,
they typically make inefficient use of the collected observations and require extensive experience to
achieve good performance.
5.2 Relational Reinforcement Learning
This section reviews Relational Reinforcement Learning (RRL) (Dzeroski et al., 2001; Otterlo, 2009)
given that RRL uses a knowledge representation similar to one used in learning for AP. The section
focuses on pure model-free RRL since techniques for learning relational action models were already
discussed in Section 3. Further explanations on model-based RRL can be found in (Boutilier et al., 2001;
Wang et al., 2007; Sanner and Kersting, 2010) also referenced as Relational, Symbolic or First Order
Dynamic Programming.
1. Knowledge representation. The learned knowledge is represented in the form of a First Order policy
π:SAthat maps relational states– coded in predicate logic– into the preferred action to apply next
for achieving a set of First Order goals. Unlike generalized policies for AP (Section 4), RRL policies
do not include information on different goals. RRL policies only capture how to solve a particular
task or a particular set of tasks, like the same task with different number of objects. RRL requires
consequently learning from scratch or at least, some kind of transfer learning (Taylor and Stone,
2007; Mehta et al., 2008) when the nature of goals change. Transfer learning is defined as exploiting
data from one or more source tasks to improve learning performance in a different target task.
AFirst Order policy can be implicitly represented as a First Order Q-function Q(< S, A >) = Rthat
maps pairs state-action SAinto their expected reward R. This representation is closely related to
the concept of heuristic function used in AP that maps states into an estimate of the cost to pay
(negative reward) for achieving the goals. In AP this cost estimate can be automatically extracted
from the problem representation by ignoring the delete effects of actions.
2. Learning examples. Learning examples consist of tuples < si, ai, si+1, ri>where aiis the action
applied at state si,si+1 is the resulting state and riis the obtained reward. Learning examples in RRL
are typically generated from random exploration.
A Review of Machine Learning for Automated Planning 27
3. Learning algorithm. There are different approaches for model-free RRL:
Relational learning of the q-value function. This approach uses relational regression to generalize
the value function. Figure 18 shows a relational regression tree that captures the q-values
of action move(Block,Block) when solving the set of tasks on(Block,Block) in a three blocks
goal_on(A,B), move(D,E).
+--yes: [0]
+--no: clear(A)?
+--yes: clear(B)?
| +--yes:[1]
| +--no:[0.9]
Figure 18 Relational regression tree for the goals on(X,Y) in a three blocks Blocksworkd.
Relational learning of the optimal policy. An alternative approach is to directly learn the policy.
This approach requires relational classification instead of relational regression. The advantage
of this approach is that, usually, it is easier to understand policies than to understand value
functions for structured domains. Figure 19 shows a RRL policy that captures when the selection
of action move(Block,Block) is optimal for solving the set of tasks on(Block,Block) in a three
blocks Blocksworld domain.
+--yes: equal(E,B)?
| +--yes: nonoptimal
| +--no: optimal
+--no: above(D,B)?
+--yes: optimal
+--no: move(A,B)?
+--yes: optimal
+--no: nonoptimal
Figure 19 RRL policy for the goals on(X,Y) in a three blocks Blocksworkd.
5.3 Implementations
With regard to learning a relational Q-function, different relational regression algorithms have been tried:
Relational regression trees (Dzeroski et al., 2001). For each pair (action, relational-goal), a regression
tree is built from a set of examples with the form (state,q-value). Leaf nodes of the tree represent
the predictions of the q-value. Test nodes of the tree represent the facts that have to hold for the
predictions to come true.
Instance based algorithms with relational distance (Driessens and Ramon, 2003). This work imple-
ments a k-nearest neighbor prediction. The prediction computes a weighted average for the q-values
of the collected examples. Weights used for the prediction are inversely proportional to the distance
to the examples. The used distance copes with relational representations for states and actions.
Relational Kernels methods (Gartner et al., 2003). These methods use the incrementally learnable
Gaussian processes for Bayesian-regression to build mappings of (state,action) pairs into q-values.
Graph kernels are used as the covariance function between state-action pairs to employ Gaussian
processes in the relational setting.
28 S.JIM´
With regard to directly learning a relational policy, (Dzeroski et al., 2001) uses relational decision trees
to determine which actions are optimal in the different environment’s states when achieving a particular
5.4 Discussion
The aims of RL are closely related to the aims of learning for AP. However, RL typically presents two
kinds of shortcomings when addressing AP problems:
Scalability. The space state of symbolic planning problems is normally huge. This space grows
exponentially with the number of objects and predicates in the problem. Many RL algorithms, like
Q-learning, require a table with one entry for each state in the state space limiting their applicability
to AP problems. A solution to this limitation is using relational models which adopt the same state
representation as symbolic planning, like done in RRL.
Generalization. RL focuses learning on the achievement of particular goals. Each time goals change
RL agents need to learn from scratch, or a transfer learning process at least. This is not exactly
the case of RRL. Given that RRL represent goals using first order predicates, RRL agents can act in
tasks with additional objects without reformulating their learned policies, although additional training
may be needed to achieve optimal (or even acceptable) performance levels. Even more, in the case
of model-based RL, the agent learns a model of the environment and exploits it to learn policies
more efficiently. The learned model is typically used to generate advise on how to explore or to plan
trajectories so that the agent can obtain higher rewards. When model-based techniques are applied
to RRL (Croonenborghs et al., 2007b) they produce symbolic action models similar to the ones
learned in AP. In this case, the main difference comes from the fact that AP learns action models
in standard planning representation languages and that the problem solving is carried out by off-the-
shelf planners.
Despite the advances in RRL, applying RRL to planning problems is still an open issue. In complex
planning tasks like the Blocksworld ones, RRL agents spend long time exploring actions without learning
anything because no rewards (goal states) are encountered. By the time being the limitations of random
exploration in RRL are relieved by two approaches. The first one uses traces of human-defined reasonable
policies to provide the learning with some positive rewards (Driessens and Matwin, 2004). The second
one applies transfer learning to the relational setting (Croonenborghs et al., 2007a).
To sum up, RRL focuses on learning policies for particular goals, like on(X,Y) for the Blocksworld,
and fail to solve problems in which interacting goals have to be achieved, for instance building a tower
of specific blocks on(X,Y), on(Y,Z), on(Z,W). In this kind of problems, traditionally addressed in AP, the
achievement of a particular goal may undo previously satisfied goals. This kind of goals must be attained
in a specific order (as happens in the Sussman’s Anomaly). As far as we know, none of the reported RRL
approaches have mechanisms to automatically capture this knowledge about the interaction of goals.
6 Conclusions
This section summarizes the contents of the paper and discusses some issues identified in the review that
are still open.
6.1 Summary
Before the mid 90’s ML was used exhaustively in AP to learn search control knowledge that improved the
performance of planners. During this period planners implemented uninformed search algorithms, so ML
helped planners achieve better performance in many domains. In the late 90’s the planning community’s
interest in ML waned for two reasons. The appearance of powerful domain-independent heuristics boosted
the performance of planners. Suddenly, the baseline for evaluating the benefit of ML shifted dramatically.
That said, existing relational learning algorithms were inefficient and performed poorly in a diversity of
A Review of Machine Learning for Automated Planning 29
Today encouraged by the applications of AP to real-world problems and the maturity of relational
learning, there is a renewed interest in learning for planning. In 2005 the International Competition on
Knowledge Engineering for Planning Systems (ICKEPS) began and in 2008 a learning track was opened
at IPC. Workshops on planning and learning have taken place periodically at the International Conference
on Automated Planning and Scheduling. ML seems to be once again one of the solutions to the challenges
of AP.
In the paper we focused on two of AP’s challenges: defining effective AP action models and defining
search control knowledge to improve the planners performance. For the first challenge, we reviewed
systems that learned action models with diverse forms of preconditions and effects. These systems
generally present learning algorithms that are effective when the appropriate learning examples are
collected although it is still not clear how to automatically collect good sets of learning examples in
AP. When learning results in imperfect action models, there are few mechanisms for diagnosing the flaws
of the models or algorithms to robustly plan with them.
For the second challenge, we reviewed different forms of search control knowledge to improve off-the-
shelf planners. This kind of knowledge has been shown to boost the performance of planners in particular
domains. Learning effective search control knowledge over a collection of domains is still challenging
since different planning domains may present very different structures. We also reviewed learning search
control for HTN planners which is a more expressive form of control knowledge that uses a different
planning paradigm based on domain-specific problem decomposition. In this planning paradigm, search
control knowledge and action models are not separated in the domain theory’s definition.
The paper also reviewed techniques for RRL, a form of RL that, like AP, uses predicate logic to
represent states and actions. Though current RRL techniques can solve relational tasks, they are focused
on achieving a particular set of goals and present generalization limitations in comparison with learning
for AP techniques. Furthermore, they have problems collecting significant experience for complex tasks
like the ones traditionally addressed in AP, where achieving a particular goal may undo previously satisfied
6.2 Open Issues
Below, we provide with a list of what we consider to be open issues or future avenues in learning for
AP. We divide these issues based on the paper’s main four topics: knowledge representation, learning
examples, learning algorithms, and exploitation of the learned knowledge:
Knowledge representation. Finding an effective knowledge representation for AP over a collection
of domains is still an open issue. If the chosen representation is not able to express key knowledge
in a given domain, the learning algorithms will not generate effective AP knowledge for this domain.
Examples of this key knowledge are the concepts of well-placed block in the Blocksworld domain or
the in-final-city concept in the Logistics domain (Khardon, 1999). Further studies need to induce what
the key knowledge for a given domain would be and what the suitable representation for expressing
it would be. In this regard, the work presented in (Yoon et al., 2008) used a set of testing problems
to evaluate the utility of the different features of the planning context when learning a generalized
policy. Another potentially relevant aspect is how the representation language affects the ability to
learn effective knowledge. In (Martin and Geffner, 2000; Cresswell et al., 2009) the authors showed
the impact of learning using object centered representations for AP but the use of new standard
representations such as SAS+ (Helmert, 2009) instead of PDDL has not yet been studied in detail.
More expressive representation formalisms like programs, temporal formulae or hierarchical repre-
sentations have been shown to describe abstract structures such as loops or hierarchies that represent
effective knowledge for many AP domains. Most of the methods for learning these representations
however need extra labelling such as annotations to mark abstract tasks, the beginning and end
of loops, end-conditions, etc. This kind of labelling cannot easily be obtained automatically from
observed executions. This knowledge cannot be directly used by off-the-shelf planners and requires
30 S.JIM´
“ad-hoc” planning systems able to exploit this kind of knowledge. Further studies in learning and in
compiling it for off-the-shelf planners are needed.
Learning examples. Implementing mechanisms to autonomously collect learning examples for AP
is still an open issue. Random explorations typically undersample the AP state and action spaces.
AP actions present preconditions that are only satisfied by specific sequences of actions which have
a low probability of being chosen by chance. Learning examples can be extracted from solutions
to AP training problems. Traditionally, these training problems are obtained by random generators
with some parameters to tune the problems difficulty. This approach has two limitations. (1) It is not
trivial to guarantee AP problems’ solvability. Showing that from a given initial state, one can reach
the goals is as hard as solving the original AP task. (2) The AP problem generators’ parameters are
domain-specific. The number and type of the world objects are different for different AP domains
and tuning these parameters to generate good quality learning examples implies domain expertise.
In (Fern et al., 2004) Random Walks were introduced to not have to have domain expertise to
obtain good learning examples. Random Walks is an automatic and domain-independent strategy for
automatically generating problems of gradually increasing walk length. In domains where complexity
depends on the number of world objects, however, this approach is not enough. An alternative
approach consists of generating training problems using an active learning approach starting from
a previous state (Fuentetaja and Borrajo, 2006). In this case the learning process has the bias imposed
by the above mentioned problem, but it can generate a random exploration of problems of increasing
difficulty in domains where complexity depends on the number of world objects. This question is
still open given that when the domain is asymmetrical, the latter approach could generate unsolvable
problems. Further studies have to be undertaken to come up with more general solutions.
Learning algorithms. The application and combination of new ML algorithms for AP is still an open
issue. ML is constantly producing new learning algorithms with potential applications for AP. In the
case of relational learning AP has only benefited from algorithms for classification and regression
but existing relational learning algorithms cover almost all machine learning tasks. Algorithms for
relational clustering or for inferring association rules (Raedt, 2008) have still not been explored in
AP. ML algorithms for propositional data can be adapted to the relational setting used in AP (Mour ˜
et al., 2010). Another example of applying new ML algorithms to AP is the use of kernel functions
to effectively match planning instances (Serina, 2010). Despite all this new ML technology, learning
algorithms present biases that affect the learning process. Research in ML for AP can take up new
research directions to study the combination of different learning algorithms to reduce the negative
effect of a given algorithm bias (Kittler, 1998).
Exploitation of the learned knowledge. These open issues in learning for AP frequently result in
imperfect learned knowledge. As the IPC-2008 learning track showed, directly applying defective
learned knowledge may damage the AP process performance. The greedier the application of the
learned knowledge, the more sensitive the AP algorithms are to the flaws of the learned knowledge.
There is a recent line of work on the development of planning and learning systems that aims to build
mechanisms to guarantee robust planning even when harmful knowledge is learned. These kind of
systems do not only focus on the learning process but also on a robust application of the learned
knowledge. Integrating learned knowledge with domain-independent heuristics generated a planning
system that was competitive with the state-of-the-art planners in a large variety of domains (Yoon
et al., 2007, 2008). Another interesting line of research is into the development of mechanisms for
evaluating the quality of the learned domain control knowledge. Such mechanisms could be useful
for defining different exploitation strategies. The only evaluation mechanisms currently available
for the learned knowledge are empirical and consist of using the learned knowledge to solve a
set of training examples to estimate its performance. Apart from the competition, another way
of exploiting learned knowledge in AP is through interaction with the final user. Mixed initiative
planning systems (Ferguson et al., 1996; Bresina et al., 2005; Cortellessa and Cesta, 2006) propose
knowledge that can be verified and modified by the user before being applied.
A Review of Machine Learning for Automated Planning 31
This review focused on learning two inputs of the AP process, the action model and knowledge for
search control. There is a full set of techniques that aim to capture information about the third input to the
AP process, the problem. Plan recognition techniques (Charniak and Goldman, 1993; Bui et al., 2002) try
to infer an agent’s goals and plans by observing the agent’s actions. The plan recognition task traditionally
assumes there is a library with the space of possible plans to be recognized but recent works do not require
the use of this library. These works (Ram’ırez and Geffner, 2009; Ram´
ırez and Geffner, 2010) can also
compute the goals that account for the observed examples using slightly modified planning algorithms.
Further studies are needed to analyze the relationship of plan recognition techniques with the learning
techniques for AP reviewed in this paper.
Aler, R., Borrajo, D., and Isasi, P. (2002). Using genetic programming to learn and improve control
knowledge. Artificial Intelligence, 141(1-2):29–56.
Amir, E. and Chang, A. (2008). Learning partially observable deterministic action models. Journal of
Artificial Intelligence Research, 33:349–402.
Bacchus, F. and Kabanza, F. (2000). Using temporal logics to express search control knowledge for
planning. Artificial Intelligence, 116(1-2):123–191.
Barto, A. and Duff, M. (1994). Monte carlo matrix inversion and reinforcement learning. In Advances in
Neural Information Processing Systems 6, pages 687–694.
Bellingham, J. and Rajan, K. (2007). Robotics in remote and hostile environments. Science,
Bellman, R. and Kalaba, R. (1965). Dynamic Programming and Modern Control Theory. Academic
Benson, S. S. (1997). Learning Action Models for Reactive Autonomous Agents. PhD thesis, Stanford
Bergmann, R. and Wilke, W. (1996). PARIS: Flexible plan adaptation by abstraction and refinement. In
Workshop on Adaptation in Case-Based Reasoning. ECAI-96.
Bertsekas, D. P. (1995). Dynamic Programming and Optimal Control. Athena Scientific.
Bertsekas, D. P. and Tsitsiklis, J. N. (1996). Neuro-Dynamic Programming (Optimization and Neural
Computation Series, 3). Athena Scientific.
Blockeel, H. and De Raedt, L. (1998). Top-down induction of first-order logical decision trees. Artificial
Intelligence, 101:285–297.
Blockeel, H., Raedt, L. D., and Ramong, J. (1998). Top-down induction of clustering trees. In In
Proceedings of the 15th International Conference on Machine Learning.
Blum, A. and Furst, M. (1995). Fast planning through planning graph analysis. In International Joint
Conference on Artificial Intelligence, IJCAI-95.
Bonet, B. and Geffner, H. (2001). Planning as heuristic search. Artificial Intelligence, 129(1–2):5–33.
Borrajo, D. and Veloso, M. (1997). Lazy incremental learning of control knowledge for efficiently
obtaining quality plans. AI Review Journal. Special Issue on Lazy Learning, 11(1-5):371–405.
Botea, A., Enzenberger, M., Mller, M., and Schaeffer, J. (2005a). Macro-FF: Improving AI Planning with
Automatically Learned Macro-Operators. Journal of Artificial Intelligence Research, 24:581–621.
32 S.JIM´
Botea, A., M¨
uller, M., and Schaeffer, J. (2007). Fast Planning with Iterative Macros. In Proceedings of
the International Joint Conference on Artificial Intelligence IJCAI-07, pages 1828–1833.
Botea, A., Mller, M., and Schaeffer, J. (2005b). Learning partial-order macros from solutions. In ICAPS05,
pages 231–240.
Boutilier, C., Reiter, R., and Price, B. (2001). Symbolic dynamic programming for first-order MDPs. In
International Joint Conference on Artificial Intelligence.
Brazdil, P., Giraud-Carrier, C., Soares, C., and Vilalta, R. (2009). Metalearning: Applications to Data
Mining. Cognitive Technologies. Springer.
Bresina, J. L., Jnsson, A. K., Morris, P. H., and Rajan, K. (2005). Mixed-initiative activity planning for
mars rovers. In IJCAI, pages 1709–1710.
Bui, H. H., Venkatesh, S., and West, G. (2002). Policy recognition in the abstract hidden markov model.
Journal of Artificial Intelligence Research, 17:2002.
Bulitko, V. and Lee, G. (2006). Learning in real-time search: A unifying framework. Journal of Artificial
Intelligence Research, 25:119–157.
Bylander, T. (1991). Complexity results for planning. In International Joint Conference on Artificial
Intelligence. IJCAI-91, Sydney, Australia.
Bylander, T. (1994). The computational complexity of propositional STRIPS planning. Artificial
Intelligence, 69(1-2):165–204.
Castillo, L., Fdez.-Olivares, J., Garc´
erez, O., and Palao, F. (2006). Bringing users and planning
technology together. experiences in SIADEX. In International Conference on Automated Planning
and Scheduling (ICAPS 2006).
Charniak, E. and Goldman, R. P. (1993). A bayesian model of plan recognition. Artif. Intell., 64(1):53–79.
Cohen, W. W. (1990). Learning approximate control rules of high utility. In International Conference on
Machine Learning.
Coles, A. and Smith, A. (2007). Marvin: A heuristic search planner with online macro-action learning.
Journal of Artificial Intelligence Research, 28:119–156.
Cortellessa, G. and Cesta, A. (2006). Evaluating Mixed-Initiative Systems: An Experimental Approach.
In Proceedings of the 16th International Conference on Automated Planning & Scheduling, ICAPS-06.
Cresswell, S., McCluskey, T. L., and West, M. (2009). Acquisition of object-centred domain models from
planning examples. In Proceedings of the 19th International Conference on Automated Planning and
Scheduling (ICAPS-09).
Croonenborghs, T., Driessens, K., and Bruynooghe, M. (2007a). Learning relational options for inductive
transfer in relational reinforcement learning. In Proceedings of the Seventeenth Conference on Inductive
Logic Programming.
Croonenborghs, T., Ramon, J., Blockeel, H., and Bruynooghe, M. (2007b). Online learning and exploiting
relational models in reinforcement learning. In Proceedings of the 20th International Joint Conference
on Artificial Intelligence, pages 726–731. AAAI press.
Cussens, J. (2001). Parameter estimation in stochastic logic programs. Machine Learning, 44(3):245–271.
Dawson, C. and Silklossly, L. (1977). The role of preprocessing in problem solving system. In
International Joint Conference on Artificial Intelligence, IJCAI-77, pages 465–471.
A Review of Machine Learning for Automated Planning 33
de la Rosa, T., Garc´
ıa-Olaya, A., and Borrajo, D. (2007). Using cases utility for heuristic planning
improvement. In International Conference on Case-Based Reasoning.
de la Rosa, T., Jim´
enez, S., and Borrajo, D. (2008). Learning relational decision trees for guiding heuristic
planning. In International Conference on Automated Planning and Scheduling (ICAPS 08).
de la Rosa, T., Jim´
enez, S., Garc´
an, R., Fern´
andez, F., Garc´
ıa-Olaya, A., and Borrajo, D. (2009).
Three relational learning approaches for lookahead heuristic planning. In Working notes of the
ICAPS’09 Workshop on Planning and Learning.
Driessens, K. and Matwin, S. (2004). Integrating guidance into relational reinforcement learning. Machine
Learning, 57:271–304.
Driessens, K. and Ramon, J. (2003). Relational instance based regression for relational reinforcement
learning. International Conference on Machine Learning.
Dzeroski, S., Raedt, L. D., and Driessens, K. (2001). Relational reinforcement learning. Machine
Learning, 43:7–52.
Edelkamp, S. (2002). Symbolic pattern databases in heuristic search planning. In International
Conference on Automated Planning and Scheduling.
Ernst, G. W. and Newell, A. (1969). GPS: A Case Study in Generality and Problem Solving. ACM
Monograph Series. Academic Press, New York, NY.
Erol, K., Nau, D. S., and Subrahmanian, V. S. (1992). On the complexity of domain-independent planning.
Artificial Intelligence, 56:223 – 254.
Estlin, T. A. and Mooney, R. J. (1996). Hybrid learning of search control for partial-order planning. In In
New Directions in AI Planning, pages 115–128. IOS Press.
Etzioni, O. (1993). Acquiring search-control knowledge via static analysis. Artificial Intelligence,
Ferguson, G., Allen, J. F., and Miller, B. (1996). Trains-95: Towards a mixed-initiative planning assistant.
In International Conference on Artificial Intelligence Planning Systems, AIPS96, pages 70–77. AAAI
Fern, A., Yoon, S., and Givan, R. (2004). Learning domain-specific control knowledge from random
walks. In International Conference on Automated Planning and Scheduling., pages 191–199.
Fern, A., Yoon, S. W., and Givan, R. (2006). Approximate policy iteration with a policy language bias:
Solving relational markov decision processes. Journal of Artificial Intelligence Research, 25:75–118.
Fikes, R., Hart, P., and Nilsson, N. J. (1972). Learning and executing generalized robot plans. Artificial
Intelligence, 3:251–288.
Fikes, R. and Nilsson, N. J. (1971). STRIPS: A new approach to the application of theorem proving to
problem solving. Artificial Intelligence, 2:189–208.
Florez, J. E., Garca, J., Torralba, A., Linares, C., Garca-Olaya, A., and Borrajo, D. (2010). Timiplan: An
application to solve multimodal transportation problems. In Proceedings of SPARK, Scheduling and
Planning Applications woRKshop, ICAPS’10.
Fox, M. and Long, D. (2003). PDDL2.1: An extension to PDDL for expressing temporal planning
domains. Journal of Artificial Intelligence Research, pages 61–124.
Fuentetaja, R. and Borrajo, D. (2006). Improving control-knowledge acquisition for planning by active
learning. In European Conference on Learning, pages 138–149.
34 S.JIM´
an, R., Fern´
andez, F., and Borrajo, D. (2006). Combining macro-operators with control
knowledge. In ILP.
an, R., Fern´
andez, F., and Borrajo, D. (In Press). A prototype-based method for classification
with time constraints: A case study on automated planning. Pattern Analysis and Applications Journal.
ınez, R. and Borrajo, D. (2000). An integrated approach of learning, planning, and execution.
Journal of Intelligent and Robotics Systems, 29:47–78.
Gartner, T., Driessens, K., and Ramon, J. (2003). Graph kernels and gaussian processes for relational
reinforcement learning. In International Conference on Inductive Logic Programming, ILP 2003.
Gerevini, A., Saetti, A., and Vallati, M. (2009a). An automatically configurable portfolio-based planner
with macro-actions: PbP. In Proceedings of the 19th International Conference on Automated Planning
and Scheduling (ICAPS-09).
Gerevini, A. E., Haslum, P., Long, D., Saetti, A., and Dimopoulos, Y. (2009b). Deterministic planning
in the fifth international planning competition: Pddl3 and experimental evaluation of the planners.
Artificial Intelligence, 173(5-6):619–668.
Ghallab, M., Nau, D., and Traverso, P. (2004). Automated Planning Theory and Practice. Morgan
Gil, Y. (1992). Acquiring Domain Knowledge for Planning by Experimentation. PhD thesis, School of
Computer Science, Carnegie Mellon University, Pittsburgh.
Gretton, C. and Thi´
ebaux, S. (2004). Exploiting first-order regression in inductive policy selection. In
Conference on Uncertainty in Artificial Intelligence.
Helmert, M. (2009). Concise finite-domain representations for pddl planning tasks. Artificial Intelligence.
andez, C. and Meseguer, P. (2007). Improving LRTA*(k). In International Joint Conference on
Artificial Intelligence, IJCAI-07, pages 2312–2317.
Hoffmann, J. and Nebel, B. (2001a). The FF planning system: Fast plan generation through heuristic
search. Journal of Artificial Intelligence Research, 14:253–302.
Hoffmann, J. and Nebel, B. (2001b). The FF planning system: Fast plan generation through heuristic
search. JAIR, 14:253–302.
Hogg, C., Kuter, U., and Mu˜
noz-Avila, H. (2009). Learning hierarchical task networks for nondetermin-
istic planning domains. In International Joint Conference on Artificial Intelligence. IJCAI-09.
Hogg, C., Mu˜
noz-Avila, H., and Kuter, U. (2008). HTN-MAKER: Learning HTNs with minimal addi-
tional knowledge engineering required. In National Conference on Artificial Intelligence (AAAI’2008).
Howe, A. E., Dahlman, E., Hansen, C., Scheetz, M., and Mayrhauser, A. V. (1999). Exploiting competitive
planner performance. In In Proceedings of the Fifth European Conference on Planning.
Ilghami, O., Mu noz-Avila, H., Nau, D. S., and Aha, D. W. (2005). Learning approximate preconditions
for methods in hierarchical plans. In International Conference on Machine learning.
Ilghami, O., Nau, D. S., and Mu˜
noz-Avila, H. (2002). CaMeL: Learning Method Preconditions for HTN
Planning. In Proceedings of the Sixth International Conference on AI Planning and Scheduling, pages
131–141. AAAI Press.
Ilghami, O., Nau, D. S., and Mu˜
noz-Avila, H. (2006). Learning to do HTN planning. In International
Conference on Automated Planning and Scheduling, ICAPS 2006.
A Review of Machine Learning for Automated Planning 35
Jaeger, M. (1997). Relational bayesian networks. In Conference on Uncertainty in Artificial Intelligence.
Jess Lanchas, Sergio Jimnez, F. F. and Borrajo, D. (2007). Learning action durations from executions. In
Working notes of the ICAPS’07 Workshop on AI Planning and Learning.
enez, S., Fern´
andez, F., and Borrajo, D. (2008). The PELA architecture: integrating planning and
learning to improve execution. In Proceedings of the Twenty-Third AAAI Conference on Artificial
Intelligence (AAAI-08). Chicago, Illinois, USA.
Kaelbling, L. P., Littman, M. L., and Moore, A. P. (1996). Reinforcement learning: A survey. Journal of
Artificial Intelligence Research, 4:237–285.
Kambhampati, S. (2007). Model-lite planning for the web age masses: The challenges of planning with
incomplete and evolving domain models. In Senior Member track of the AAAI, Seattle, Washington,
USA. AAAI Press/MIT Press.
Kambhampati, S. and Hendler, J. A. (1992). A validation structure-based theory of plan modification and
reuse. Artificial Intelligence Journal, 55:193–258.
Keller, R. (1987). The Role of Explicit Contextual Knowledge in Learning Concepts to Improve
Performance. PhD thesis, Rutgers University.
Kersting, K. and Raedt, L. D. (2001). Towards combining inductive logic programming with Bayesian
networks. In International Conference on Inductive Logic Programming, pages 118–131.
Khardon, R. (1999). Learning action strategies for planning domains. Artificial Intelligence, 113:125–148.
Kittler, J. (1998). Combining classifiers: A theoretical framework. Number 1.
Korf, R. E. (1985). Macro-operators: A weak method for learning. Artificial Intelligence, 1985, 26:35–77.
Korf, R. E. (1990). Real-time heuristic search. Artificial Intelligence, 42(2-3):189–211.
Larkin, J., Reif, F., and Carbonell, J. (1986). Fermi: A flexible expert reasoner with multi-domain
inference. Cognitive Science, 9.
Leckie, C. and Zukerman, I. (1991). Learning search control rules for planning: An inductive approach.
In International Workshop on Machine Learning, pages 422–426, Evanston, IL. Morgan Kaufmann.
Martin, M. and Geffner, H. (2000). Learning generalized policies in planning using concept languages.
In International Conference on Artificial Intelligence Planning Systems, AIPS00.
Mcallester, D. and Givan, R. (1989). Taxonomic syntax for first order inference. Journal of the ACM,
McCluskey, T. L. (1987). Combining weak learning heuristics in general problem solvers. In IJCAI’87:
Proceedings of the 10th international joint conference on Artificial intelligence, pages 331–333, San
Francisco, CA, USA. Morgan Kaufmann Publishers Inc.
McGann, C., Py, F., Rajan, K., Ryan, J., and Henthorn, R. (2008). Adaptive control for autonomous
underwater vehicles. In National Conference on Artificial Intelligence (AAAI’2008).
Mehta, N., Natarajan, S., Tadepalli, P., and Fern, A. (2008). Transfer in variable-reward hierarchical
reinforcement learning. Machine Learning, 73(3):289–312.
Minton, S. (1988). Learning Effective Search Control Knowledge: An Explanation-Based Approach.
Kluwer Academic Publishers, Boston, MA.
Mitchell, T., Utgoff, T., and Banerji, R. (1982). Machine Learning: An Artificial Intelligence Approach,
chapter Learning problem solving heuristics by experimentation. Morgan Kaufmann.
36 S.JIM´
Mitchell, T. M. (1997). Machine Learning. McGraw-Hill, New York.
ao, K., Petrick, R. P. A., and Steedman, M. (2008). Using kernel perceptrons to learn action effects
for planning. In Proceedings of the International Conference on Cognitive Systems (CogSys 2008).
ao, K., Petrick, R. P. A., and Steedman, M. (2010). Learning action effects in partially observable
domains. In European Conference on Artificial Intelligence.
Muggleton, S. (1995). Stochastic logic programs. In International Workshop on Inductive Logic
Muise, C., McIlraith, S., Baier, J. A., and Reimer, M. (2009). Exploiting N-gram analysis to predict
operator sequences. In 19th International Conference on Automated Planning and Scheduling (ICAPS).
noz-Avila, Aha, H., Breslow, D., and Nau, L. (1999). HICAP: An interactive case based planning
architecture and its application to noncombatant evacuation operations. In Conference on Innovative
Applications of Artificial Intelligence. IAAI-99.
Nau, D., Ilghami, O., Kuter, U., Murdock, J. W., Wu, D., and Yaman, F. (2003). Shop2: An htn planning
system. Journal of Artificial Intelligence Research, 20:379–404.
Nau, D. S., Smith, S. J., and Erol, K. (1998). Control strategies in htn planning: Theory versus practice.
In In AAAI-98/IAAI-98 Proceedings.
Nayak, P., Kurien, J., Dorais, G., Millar, W., Rajan, K., and Kanefsky, R. (1999). Validating the ds-1
remote agent experiment. In Artificial Intelligence, Robotics and Automation in Space.
Newton, M. A. H., Levine, J., Fox, M., and Long, D. (2007). Learning macro-actions for arbitrary planners
and domains. In International Conference on Automated Planning and Scheduling.
Nilsson, N. J. (1984). Shakey the robot. Technical Report 323, AI Center, SRI International, Menlo Park,
Oates, T. and Cohen, P. R. (1996). Searching for planning operators with context-dependent and
probabilistic effects. In National Conference on Artificial Intelligence.
Otterlo, M. V. (2009). The Logic of Adaptive Behavior: Knowledge Representation and Algorithms for
Adaptive Sequential Decision Making under Uncertainty in First-Order and Relational Domains. IOS
Pasula, H. M., Zettlemoyer, L. S., and Kaelbling, L. P. (2007). Learning symbolic models of stochastic
domains. Journal of Artificial Intelligence Research, 29:309–352.
Porteous, J. and Sebastia, L. (2004). Ordered landmarks in planning. Journal of Artificial Intelligence
Research, 22:215–278.
Quinlan, J. and Cameron-Jones, R. (1995). Introduction of logic programs: FOIL and related systems.
New Generation Computing, Special issue on Inductive Logic Programming.
Raedt, L. D. (2008). Logical and Relational Learning. Springer, Berlin Heidelberg.
Ram’ırez, M. and Geffner, H. (2009). Plan recognition as planning. In IJCAI’09: Proceedings of the 21st
international jont conference on Artifical intelligence.
ırez, M. and Geffner, H. (2010). Probabilistic plan recognition using off-the-shelf classical planners.
In National Conference on Artificial Intelligence (AAAI’2010).
Reynolds, S. I. (2002). Reinforcement Learning with Exploration. PhD thesis, The University of
Birmingham, UK.
A Review of Machine Learning for Automated Planning 37
Richardson, M. and Domingos, P. (2006). Markov logic networks. Machine Learning, 62:107–136.
Rivest, R. L. (1987). Learning decision lists. Machine Learning, 2(3):229–246.
Sanner, S. and Boutilier, C. (2005). Approximate linear programming for first-order mdps. In In
Proceedings of the 21st Conference on Uncertainty in Artificial Intelligence, pages 509–517.
Sanner, S. and Boutilier, C. (2006). Practical linear value-approximation techniques for first-order mdps.
In Proceedings of the 22nd Conference in Uncertainty in Artificial Intelligence.
Sanner, S. and Kersting, K. (2010). Symbolic dynamic programming for first–order pomdps. In M. Fox,
D. P., editor, Twenty–Fourth AAAI Conference on Artificial Intelligence (AAAI–10), Atlanta, USA.
AAAI Press.
Serina, I. (2010). Kernel functions for case-based planning. Artif. Intell., 174(16-17):1369–1406.
Shavlik, J. W. (1989). Acquiring recursive and iterative concepts with explanation-based learning. In
Machine Learning.
Shen, W. and Simon (1989). Rule creation and rule learning through environmental exploration. In
International Joint Conference on Artificial Intelligence, IJCAI-89, pages 675–680.
Srivastava, S., Immerman, N., and Zilberstein, S. (2008). Learning generalized plans using abstract
counting. In National Conference on Artificial Intelligence (AAAI’2008).
Strehl, A. L. and Littman, M. L. (2005). A theoretical analysis of model-based interval estimation. In
Proceedings of the Twenty-second International Conference on Machine Learning (ICML-05, pages
Sutton, R. S. and Barto, A. G. (1998). Reinforcement Learning: An Introduction (Adaptive Computation
and Machine Learning). The MIT Press.
Taylor, M. E. and Stone, P. (2007). Cross-domain transfer for reinforcement learning. In International
Conference on Machine Learning. ICML.
Theocharous, G. and Kaelbling, L. P. (2003). Approximate planning in POMDPs with macro-actions. In
In Proceedings of Advances in Neural Information Processing Systems 16.
ebaux, S., Hoffmann, J., and Nebel, B. (2005). In defense of pddl axioms. Artificial Intelligence,
Veloso, M. M. and Carbonell, J. G. (1993). Derivational analogy in prodigy: Automating case acquisition,
storage, and utilization. Machine Learning, 10:249–278.
Veloso, M. M., P ´
erez, M. A., and Carbonell, J. G. (1990). Nonlinear planning with parallel resource
allocation. In In Proceedings of the DARPA Workshop on Innovative Approaches to Planning,
Scheduling, and Control, pages 207–212. Morgan Kaufmann.
Vrakas, D., Tsoumakas, G., Bassiliades, N., and Vlahavas, I. P. (2005). HAPRC: an automatically
configurable planning system. AI Commun., 18(1):41–60.
Walsh, T. J. and Littman, M. L. (2008). Efficient learning of action schemas and web-service descriptions.
In AAAI’08: Proceedings of the 23rd national conference on Artificial intelligence, pages 714–719.
AAAI Press.
Wang, C., Joshi, S., and Khardon, R. (2007). First order decision diagrams for relational MDPs. In
International Joint Conference on Artificial Intelligence, IJCAI-07.
38 S.JIM´
Wang, X. (1994). Learning planning operators by observation and practice. In International Conference
on AI Planning Systems, AIPS-94.
Watkins, C. J. C. H. (1989). Learning from Delayed Rewards. PhD thesis, King’s College, Oxford.
Wiering, M. (1999). Explorations in efficient reinforcement learning. PhD thesis, University of
Amsterdam IDSIA, The Netherlands.
Winner, E. and Veloso, M. (2003). DISTILL: Towards learning domain-specific planners by example. In
International Conference on Machine Learning, ICML’03.
Xu, Y., Fern, A., and Yoon, S. W. (2007). Discriminative learning of beam-search heuristics for planning.
In International Joint Conference on Artificial Intelligence.
Yang, Q., Wu, K., and Jiang, Y. (2007). Learning action models from plan traces using weighted max-sat.
Artificial Intelligence Journal, 171:107–143.
Yoon, S., Fern, A., and Givan, R. (2002). Inductive policy selection for first-order MDPs. In Conference
on Uncertainty in Artificial Intelligence, UAI02.
Yoon, S., Fern, A., and Givan, R. (2006). Learning heuristic functions from relaxed plans. In International
Conference on Automated Planning and Scheduling (ICAPS-2006).
Yoon, S., Fern, A., and Givan, R. (2007). Using learned policies in heuristic-search planning. In
International Joint Conference on Artificial Intelligence.
Yoon, S., Fern, A., and Givan, R. (2008). Learning control knowledge for forward search planning. J.
Mach. Learn. Res., 9:683–718.
Yoon, S. and Kambhampati, S. (2007). Towards model-lite planning: A proposal for learning and planning
with incomplete domain models. In ICAPS2007 Workshop on Artificial Intelligence Planning and
Younes, H., Littman, M. L., Weissman, D., and Asmuth, J. (2005). The first probabilistic track of the
international planning competition. Journal of Artificial Intelligence Research, 24:851–887.
Zelle, J. and Mooney, R. (1993). Combining FOIL and EBG to speed-up logic programs. In International
Joint Conference on Artificial Intelligence. IJCAI-93.
Zhuo, H., Li, L., Yang, Q., and Bian, R. (2008). Learning action models with quantified conditional effects
for software requirement specification. In ICIC ’08: Proceedings of the 4th international conference
on Intelligent Computing, pages 874–881, Berlin, Heidelberg. Springer-Verlag.
Zimmerman, T. and Kambhampati, S. (2003). Learning-assisted automated planning: looking back, taking
stock, going forward. AI Magazine, 24:73 – 96.
... Fundamentally, CTF competitions are very specific scenarios which do not account for many variables in typical PT, but the study was significant for investigating the relevance of different RL techniques. An additional noteworthy proposal aimed to automate exploitation and postexploitation by combining deep RL and the PowerShell empire post-exploitation framework [52]. ...
Full-text available
Penetration Testing (PT) is an offensive method for assessing and evaluating the security of digital asset by planning, generating, and executing all or some of the possible attacks that aim to exploit its vulnerabilities. In large networks, penetration testing become repetitive, complex and resources consuming despite the use of autonomous tools. To maintain the consistency and efficiency of PT in medium and large network context. it is imperative to go through making it intelligent and optimized which will allow regular and systematic testing without having to provide a prohibitive amount of human labor in one hand and reducing the precious consumed time and tested system downtime in another hand. Reinforcement Learning (RL) led testing will unburden human experts from the heavy repetitive tasks and unveil special and complex situations such as unusual vulnerabilities or combined non-obvious combinations which are often ignored in manual testing. In this research, we are concerned with the specific context of improving current automated testing systems and making them intelligent, targeted, and efficient by embedding reinforcement learning techniques where it is relevant. The proposed Intelligent Automated Penetration Testing Framework (IAPTF) utilizes RL because of its relevance to sequential decision-making problems, it relies on a model based RL where planning and learning are combined and decomposed tasks to represent it as POMDP domain accounting for major PT features, tasks and information flowchart to realistically reflect the real-world context. The problem is then solved on an external POMDP-solver using different algorithms to identify most efficient options. As we encountered a huge scaling-up challenges in solving large POMDP which reflect the regular representation of PT on large networks, we propose thus a Hierarchical representation on which we divided large networks into security clusters and enabling IAPTF to deal with each cluster separately as small networks (intra-clusters), later we proceed to the testing of the network of clusters heads to ensure covering all possible complex and multistep attacking vectors largely adopted by nowadays hackers. The obtained results are unanimous and defeat both previous results and any human performances in term of consumed time, number tested vectors and accuracy especially in large networks. The learning is the second strength of our new model, as the generalization of the extracted knowledge become easier and allowing therefore the re-usability notably in the case of retesting the same network with few changes which is often the real-world context in PT. The performance enhancement and the knowledge extracted, and reuse confirm the efficiency, accuracy, and suitability of our proposed framework. Finally, IAPTF is designed to offload and ultimately replace human expert and to be independent, comprehensive, and versatile so it can integrate any automated PT platform or toolkit. Initially, the framework connects directly with Metasploit and Nessus APIs as both free versions coding architecture allows to perform such utilization. Keywords: penetration testing; artificial intelligence; machine learning; reinforcement learning; hierarchical reinforcement learning; network security auditing; offensive cyber-security; vulnerability Assessment
... Different goals can also be considered by coding them as part of the state representation, e.g. using static state variables. Since the early days of AI planning, hierarchies, LTL formulae, and policies, are also used to specify sketches of general solutions [46]. In the planning literature these solution sketches are often called domain-specific control knowledge, since they are traditionally used to control the planning process, and they apply to the entire set of classical planning instances that belong to a given domain [47,48,37,38]. ...
Planning as heuristic search is one of the most successful approaches to classical planning but unfortunately, it does not extend trivially to Generalized Planning (GP). GP aims to compute algorithmic solutions that are valid for a set of classical planning instances from a given domain, even if these instances differ in the number of objects, the number of state variables, their domain size, or their initial and goal configuration. The generalization requirements of GP make it impractical to perform the state-space search that is usually implemented by heuristic planners. This paper adapts the planning as heuristic search paradigm to the generalization requirements of GP, and presents the first native heuristic search approach to GP. First, the paper introduces a new pointer-based solution space for GP that is independent of the number of classical planning instances in a GP problem and the size of those instances (i.e. the number of objects, state variables and their domain sizes). Second, the paper defines a set of evaluation and heuristic functions for guiding a combinatorial search in our new GP solution space. The computation of these evaluation and heuristic functions does not require grounding states or actions in advance. Therefore our GP as heuristic search approach can handle large sets of state variables with large numerical domains, e.g.~integers. Lastly, the paper defines an upgraded version of our novel algorithm for GP called Best-First Generalized Planning (BFGP), that implements a best-first search in our pointer-based solution space, and that is guided by our evaluation/heuristic functions for GP.
... The most relevant published work is a survey of machine learning methods for automated planning [Jiménez et al., 2012] where a section is dedicated to macroactions. This survey covers only 4 studies that are also covered in this paper, published between 1977 and 2007. ...
Full-text available
Automated planning is a prominent area of Artificial Intelligence, and an important component for intelligent autonomous agents. A cornerstone of domain-independent planning is the separation between planning logic, i.e. the automated reasoning side, and the knowledge model, that encodes a formal representation of domain knowledge needed to reason upon a given problem to synthesise a solution plan. Such a separation enables the use of reformulation techniques, which transform how a model is represented in order to improve the efficiency of plan generation. Over the past decades, significant research effort has been devoted to the design of reformulation techniques. In this paper, we present a systematic review of the large body of work on reformulation techniques for classical planning, aiming to provide a holistic view of the field and to foster future research in the area. As a tangible outcome, we provide a qualitative comparison of the existing classes of techniques, that can help researchers gain an overview of their strengths and weaknesses.
... The task planning problem can be represented and solved with a multitude of formalisms and frameworks depending on the specific application. An overview of possible approaches is provided by several authors, e.g., by Ghallab et al. (2004Ghallab et al. ( , 2016 mainly in the context of planning for AI and with only some insights in generic task-motion planning for robotic applications; Jiménez et al. (2012) for learning-based planning; and Karpas and Magazzeni (2020); Nakawala et al. (2018), who considered generic action representation formalisms for robotics. The purpose of our paper is to provide a comprehensive review on the state-of-the-art solutions for representing and solving the task planning problem for deliberative robots, with a focus on logic programming tools. ...
Full-text available
Over the last decade, the use of robots in production and daily life has increased. With increasingly complex tasks and interaction in different environments including humans, robots are required a higher level of autonomy for efficient deliberation. Task planning is a key element of deliberation. It combines elementary operations into a structured plan to satisfy a prescribed goal, given specifications on the robot and the environment. In this manuscript, we present a survey on recent advances in the application of logic programming to the problem of task planning. Logic programming offers several advantages compared to other approaches, including greater expressivity and interpretability which may aid in the development of safe and reliable robots. We analyze different planners and their suitability for specific robotic applications, based on expressivity in domain representation, computational efficiency and software implementation. In this way, we support the robotic designer in choosing the best tool for his application.
... The task planning problem can be represented and solved with a multitude of formalisms and frameworks depending on the specific application. An overview of possible approaches is provided by several authors, e.g., by Ghallab et al. (2004Ghallab et al. ( , 2016 mainly in the context of planning for AI and with only some insights in generic task-motion planning for robotic applications; Jiménez et al. (2012) for learning-based planning; and Karpas and Magazzeni (2020); Nakawala et al. (2018), who considered generic action representation formalisms for robotics. The purpose of our paper is to provide a comprehensive review on the state-of-the-art solutions for representing and solving the task planning problem for deliberative robots, with a focus on logic programming tools. ...
Full-text available
Over the last decade, the use of robots in production and daily life has increased. With increasingly complex tasks and interaction in different environments including humans , robots are required a higher level of autonomy for efficient deliberation. Task planning is a key element of deliberation. It combines elementary operations into a structured plan to satisfy a prescribed goal, given specifications on the robot and the environment. In this manuscript, we present a survey on recent advances in the application of logic programming to the problem of task planning. Logic programming offers several advantages compared to other approaches, including greater expressivity and interpretability which may aid in the development of safe and reliable robots. We analyze different planners and their suitability for specific robotic applications, based on expressivity in domain representation, computational efficiency and software implementation. In this way, we support the robotic designer in choosing the best tool for his application.
... The states in the TSs are composed of Boolean propositions to denote the current planning phase. STRIPS has been successfully applied to numerous tasks (Bylander, 1994;Jiménez et al., 2012) and updated over the years (Garrett et al., 2017;Aineto et al., 2018). Yet, it is challenging to apply STRIPS to applications that involve conditional actions. ...
Full-text available
Many tasks require robots to manipulate objects while satisfying a complex interplay of spatial and temporal constraints. For instance, a table setting robot first needs to place a mug and then fill it with coffee, while satisfying spatial relations such as forks need to placed left of plates. Existing solutions are often not robot-agnostic, not expressive enough to monitor and plan complex tasks, and require users to manually define object relations. We propose the spatio-temporal framework SpaTiaL that unifies the specification, monitoring, and planning of object-centric robotic tasks. SpaTiaL captures diverse spatial relations between objects and temporal task patterns. Our experiments with recorded data, simulations, and robots demonstrate how SpaTiaL provides real-time monitoring and facilitates online planning. SpaTiaL is open source and easily expandable to new object relations and robotic applications.
Planning is a computationally expensive process, which can limit the reactivity of autonomous agents. Planning problems are usually solved in isolation, independently of similar, previously solved problems. The depth of search that a planner requires to find a solution, known as the planning horizon, is a critical factor when integrating planners into reactive agents. We consider the case of an agent repeatedly carrying out a task from different initial states. We propose a combination of classical planning and model-free reinforcement learning to reduce the planning horizon over time. Control is smoothly transferred from the planner to the model-free policy as the agent compiles the planner’s policy into a value function. Local exploration of the model-free policy allows the agent to adapt to the environment and eventually overcome model inaccuracies. We evaluate the efficacy of our framework on symbolic PDDL domains and a stochastic grid world environment and show that we are able to significantly reduce the planning horizon while improving upon model inaccuracies.KeywordsPlanningPlanning horizonReinforcement learning
We investigate artificial intelligence and machine learning methods for optimizing the adversarial behavior of agents in cybersecurity simulations. Our cybersecurity simulations integrate the modeling of agents launching Advanced Persistent Threats (APTs) with the modeling of agents using detection and mitigation mechanisms against APTs. This simulates the phenomenon of how attacks and defenses coevolve. The simulations and machine learning are used to search for optimal agent behaviors. The central question is: under what circumstances, is one training method more advantageous than another? We adapt and compare a variety of deep reinforcement learning (DRL), evolutionary strategies (ES) and Monte Carlo Tree Search methods within Connect 4, a baseline game environment, and on both a simulation supporting a simple APT threat model, SNAPT, as well as CyberBattleSim, an open-source cybersecurity simulation. Our results show that when attackers are trained by DRL and ES algorithms, as well as when they are trained with both algorithms being used in alternation, they are able to effectively choose complex exploits that thwart a defense. The algorithm that combines DRL and ES achieves the best comparative performance when attackers and defenders are simultaneously trained, rather than when each is trained against its non-learning counterpart.
Autonomous robots are increasingly served in real-world unstructured human environments with complex long-horizon tasks such as restaurant serving and office delivery. Task and motion planning (TAMP) is a recent research method in Artificial Intelligence Planning for these applications. TAMP integrates high-level abstract reasoning with the low-level geometric feasibility check and thus is more comprehensive than traditional task planning methods. While regular TAMP approaches are challenged by different types of uncertainties and the generalization of various applications when implemented in real-world scenarios. This paper systematically reviews the most relevant approaches to TAMP and classifies them according to their features and emphasis; it categorizes the challenges and presents online TAMP and machine learning-based TAMP approaches for addressing them.
This paper introduces a new representation for Boolean functions, called decision lists, and shows that they are efficiently learnable from examples. More precisely, this result is established for k-;DL – the set of decision lists with conjunctive clauses of size k at each decision. Since k-DL properly includes other well-known techniques for representing Boolean functions such as k-CNF (formulae in conjunctive normal form with at most k literals per clause), k-DNF (formulae in disjunctive normal form with at most k literals per term), and decision trees of depth k, our result strictly increases the set of functions that are known to be polynomially learnable, in the sense of Valiant (1984). Our proof is constructive: we present an algorithm that can efficiently construct an element of k-DL consistent with a given set of examples, if one exists.
In explanation-based learning, a specific problem's solution is generalized into a form that can be later used to solve conceptually similar problems. Most research in explanation-based learning involves relaxing constraints on the variables in the explanation of a specific example, rather than generalizing the graphical structure of the explanation itself. However, this precludes the acquisition of concepts where an iterative or recursive process is implicitly represented in the explanation by a fixed number of applications. This paper presents an algorithm that generalizes explanation structures and reports empirical results that demonstrate the value of acquiring recursive and iterative concepts. The BAGGER2 algorithm learns recursive and iterative concepts, integrates results from multiple examples, and extracts useful subconcepts during generalization. On problems where learning a recursive rule is not appropriate, the system produces the same result as standard explanation-based methods. Applying the learned recursive rules only requires a minor extension to a PROLOG-like problem solver, namely, the ability to explicitly call a specific rule. Empirical studies demonstrate that generalizing the structure of explanations helps avoid the recently reported negative effects of learning.
Conference Paper
I use the term logical and relational learning (LRL) to refer to the subfield of machine learning and data mining that is concerned with learning in expressive logical or relational representations. It is the union of inductive logic programming, (statistical) relational learning and multi-relational data mining and constitutes a general class of techniques and methodology for learning from structured data (such as graphs, networks, relational databases) and background knowledge. During the course of its existence, logical and relational learning has changed dramatically. Whereas early work was mainly concerned with logical issues (and even program synthesis from examples), in the 90s its focus was on the discovery of new and interpretable knowledge from structured data, often in the form of rules or patterns. Since then the range of tasks to which logical and relational learning has been applied has significantly broadened and now covers almost all machine learning problems and settings. Today, there exist logical and relational learning methods for reinforcement learning, statistical learning, distance- and kernel-based learning in addition to traditional symbolic machine learning approaches. At the same time, logical and relational learning problems are appearing everywhere. Advances in intelligent systems are enabling the generation of high-level symbolic and structured data in a wide variety of domains, including the semantic web, robotics, vision, social networks, and the life sciences, which in turn raises new challenges and opportunities for logical and relational learning. These developments have led to a new view on logical and relational learning and its role in machine learning and artificial intelligence. In this talk, I shall reflect on this view by identifying some of the lessons learned in logical and relational learning and formulating some challenges for future developments.
I analyze the computational complexity of extended propositional STRIPS planning, i.e., propositional STRIPS planning augmented with a propositional domain theory for inferring additional effects. The difficulties of formalizing the extended STRIPS assumption are finessed by requiring a preference ordering of all literals; roughly, if two literals are true of the previous state, and if it is inconsistent to assert both in the next state, then the ordering indicates which literal remains true. My primary result is that planning with definite Horn domain theories is PSPACE-complete even if operators are limited to zero preconditions and one postcondition. I also analyze the complexity of planning with Krom theories. These results in combination with previous analyses are not encouraging for domain-independent planning.