Content uploaded by The Anh Han
Author content
All content in this area was uploaded by The Anh Han on Mar 03, 2015
Content may be subject to copyright.
AUTHOR COPY
AI Communications 26 (2013) 237–246 237
DOI 10.3233/AIC-130559
IOS Press
Research Summary
State-of-the-art of intention recognition
and its use in decision making
The Anh Han a,b and Luís Moniz Pereira a,∗
aCentro de Inteligência Artificial (CENTRIA), Departamento de Informática, Faculdade de Ciências e Tecnologia,
Universidade Nova de Lisboa, Caparica, Portugal
E-mail: lmp@fct.unl.pt
bAI-lab, Vrije Universiteit Brussel, Pleinlaan 2, 1050 Brussels, Belgium
E-mail: h.anh@ai.vub.ac.be
Abstract. Intention recognition is the process of becoming aware of the intentions of other agents, inferring them through ob-
served actions or effects on the environment. Intention recognition enables pro-activeness, in cooperating or promoting coop-
eration, and in pre-empting danger. Technically, intention recognition can be performed incrementally as you go along, which
amounts to learning. Intention recognition can also use past experience from a database of past interactions, not necessarily with
the same agent. Bayesian Networks (BN) can be employed to dynamically summarize general statistical evidence, furnishing
heuristic information to link with the situation specific information, about which logical reasoning can take place, and decisions
made on actions to be performed, possibly involving actions to obtain new observations. This situated reasoning feeds into the
BN to tune it, and back again into the logic component. In this article, we provide a review bearing on the state-of-the-art work
on intention and plan recognition, which includes a comparison with our recent research, where we address a number of im-
portant issues of intention recognition. We also argue for an integrative approach to intention-based decision-making that uses a
combination of Logic Programming and Bayesian Networks.
Keywords: Intention recognition, decision making, Bayesian networks, logic programming
1. Intention recognition
We consider intention recognition in a dynamic,
real-world environment. An important aspect of inten-
tions is future-directedness, i.e., if we intend some-
thing now, we mean to execute a course of actions to
achieve something in the future [8,19,75,85]. Most ac-
tions may be executed only at a far distance in time.
During that period, the world is changing, and the ini-
tial intention may be changed to a more appropriate
one or even abandoned [9,26]. An intention recogni-
tion method should take into account these changes,
and, when necessary, be able to reevaluate the inten-
*Corresponding author: Luís Moniz Pereira, Centro de Inteligên-
cia Artificial (CENTRIA), Departamento de Informática, Faculdade
de Ciências e Tecnologia, Universidade Nova de Lisboa, 2829-516
Caparica, Portugal. E-mail: lmp@fct.unl.pt.
tion recognition model, depending on some time limit;
in addition, as new actions are observed, the model
should be reconfigurable to incorporate them. In other
words, the model should be incremental and, further-
more, the intention recognition prediction is available
at anytime.
Generally, intention recognition (also called goal
recognition) is defined as the process of becoming
aware of the intention of another agent and, more tech-
nically, as the problem of inferring an agent’s intention
through its actions and their effects on the environment
[2,17,44,88]. For the recognition task, the distinction
between my model of me and my model of another
should be mentioned, though I can use my model of me
to imagine the same model for the other. This is known
as the “Theory of Mind” theory [18,70,90], which neu-
rologically reposes in part on “mirror neurons” as sup-
porting evidence [48,58,74].
0921-7126/13/$27.50 ©2013 – IOS Press and the authors. All rights reserved
AUTHOR COPY
238 T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making
Plan recognition is closely related to intention
recognition, extending it to also recognize the plan the
observed agent is following in order to achieve his in-
tention [2,79]. Mere intention recognition is performed
in domains in which it is preferred to have a fast detec-
tion of mere user goal/intention rather than a more pre-
cise but time consuming detection of the complete user
plan, e.g., in the interface agents domain [2,46,54].
Generally, the input to both intention and plan recog-
nition systems are a set of conceivable intentions and
a set of plans achieving each intention, given in terms
of a plan library [17,26] or a plan corpus [3,6,7]. In-
tention recognition is distinct from planning, as goals
are not known a priori, and presumed goals are subject
to defeasibility. There are also generative approaches
based on planning algorithms, which do not require a
plan library/corpus (e.g., see [72]).
Intention and plan recognition have been applied
and shown to be useful in a wide range of applica-
tion domains [79], including story understanding [15],
human-computer interaction and interface-agents sys-
tems [2,45,53], traffic monitoring [71], assistive living
(e.g. Elder Care, Ambient Intelligence) [23,29,33,66,
67,76,88] and military settings [44,56].
The future-directedness of intentions also means
that, once an agent intends something, he has settled on
a particular course of action [8,19,75,85]. This makes
the intentions relatively stable, pending new informa-
tion. An agent who made the decision to act in a certain
way commits to sticking to this decision for the rea-
sons which led to it, unless counterbalancing reasons
meanwhile appear and trigger further deliberations. In
other words, intentions are relatively resistant to recon-
sideration unless there are ponderable reasons to do so
[8,9,75]. Following this, any attempt to tackle the is-
sues of intention change or abandonment cannot solely
be based on the observed actions. The reasons why the
intention is changed or abandoned must be taken into
account. The reasons can be changes in the environ-
ment (possibly made by other agents) that impel the
observed agent to refrain from following his initial in-
tention. And here the context-dependent modeling ap-
pears to be unavoidable.
Sometimes Bayesian Networks (BNs) are used as
the intention recognition model. The flexibility of BNs
for representing probabilistic dependencies and the ef-
ficiency of inference methods for BN have made them
an extremely powerful and natural tool for problem
solving under uncertainty [61,62]. The directed acyclic
graph structure of the network contains representa-
tions of both conditional dependencies and indepen-
dencies between the random variables represented by
the graph nodes. The probabilistic information is com-
pactly given by conditional probability distribution ta-
bles for each and every node. To perform intention
recognition, we constructs a three-layer BN [65,67] –
justified based on Heinze’s causal intentional model
[44,88] – and use it for evidential reasoning from ob-
servations to intention hypotheses.
One may surmise a knowledge representation
method to support incremental BN model construc-
tion for performing intention recognition during run-
time, from an initially given domain knowledge base.
As more actions are observed, a new BN is con-
structed from the previous one reinforcing some inten-
tions whilst ruling out others. This incremental method
allows domain experts to specify knowledge in terms
of small and simple BN fragments, which can be easily
maintained and changed, and which are used to com-
pose the situated ongoing BN model. Alternatively,
these fragments can be easily learned from data. More
rarely, one proposes a method to represent relation-
ships among intentions, when considering the case of
agents that may pursue multiple intentions simultane-
ously. This is an indispensable aspect, but mostly omit-
ted in prior work, which however allows to sometimes
significantly decrease the complexity of the probability
inference [27].
Some methods are generally motivated by the fact
that knowledge experts often consider a related set
of variables together, and organize domain knowledge
in larger chunks. An ability to represent conceptually
meaningful groupings of variables and their interre-
lationships facilitates both knowledge elicitation and
knowledge base maintenance [52]. To this end, there
have been several methods proposed for BN construc-
tion from small and easily maintained network frag-
ments [51,52,55,59,61,69,95]. In essence, a combina-
tion of BNs is a graph that includes all nodes and links
of the networks, where nodes with the same name are
combined into a common node. The main issue for a
combination method is how the influence of different
parents of the common node can be combined in the
new network, given the partial influence of each par-
ent in the corresponding fragment. The most exten-
sively used and popular combination method is Noisy-
Or, firstly proposed by [61] for BNs of Boolean vari-
ables, and generalized by [21,87] for the general case
of arbitrary domains.
As such, in our work [30,31,34,35], we have devel-
oped an intention recognition method that possesses
several important features: (i) The method is context-
AUTHOR COPY
T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making 239
dependent and incremental, enabling incremental con-
struction of a three-layer Bayesian Network model as
more actions are observed, and in a context-dependent
manner, which, in addition, relies on a logic program-
ming knowledge base concerning the context; (ii) The
Bayesian Network is composed from a knowledge base
of readily specified and readily maintained Bayesian
Network fragments with simple structures, thereby en-
abling an efficient acquisition of the corresponding
knowledge base (engineered either by domain experts
or else automatically from a plan corpus); and (iii) The
method addresses the issue of intention change and
abandonment, and can appropriately resolve the issue
of the recognition of multiple intentions.
To illustrate some important aspects of our method,
let us consider the following example.
Example 1.1 (Elder’s intention recognition). An el-
der stays alone in her apartment. The assistant system
(with the capability of intention recognition) observes
that she is looking for something in the living room. In
order to assist her, the system needs to figure out what
she intends to find.
The first step is to select, from a given knowledge
base of BN fragments, the fragments for action Look-
ing, which consist of a single intention connecting to
the action (Fig. 1, Box A). This process is context-
dependent, in the sense that whether an intention may
give rise to an action depends on the situation the ac-
tion is observed in. Suppose that the selected intentions
are: something to read (Book); something to drink (Wa-
ter); and the light switch (Switch). They are then com-
bined using the Noisy-OR method, resulting in a single
BN with a common action node (Fig. 1, Box A).
Next, the fragments for each intention are selected
from the given knowledge base (Fig. 1, Box B). These
fragments are then plugged into the combined net-
work in Box A, resulting in a three-layer BN in Fig. 1,
Box C, upon which intention recognition is performed.
This consists in computing the probability for each in-
tention in the current network, conditional on current
observations (e.g. the state of the light, and the ob-
served action Looking). Recall that to perform inten-
tion recognition, we construct a three-layer BN [65,
67] – justified in being based on Heinze’s causal inten-
tional model [44,88]. For further details and examples
see [30, Chapter 2], [34].
As mentioned earlier, the advantage of this dynamic
and context-dependent construction of the model is
that the obtained BN usually has a significantly re-
duced size, compared to when a large and general
network is built and used for all foreseen cases.
The context-dependent selection of the fragments can
sometimes lead to the final solution without having
to perform probabilistic inference; for instance when
there is a single conceivable intention giving rise to the
current observed actions given the situation at hand.
Fig. 1. Incremental intention recognition method via dynamic selection and construction of Bayesian Network. (Box A) (Context-dependent) se-
lection of fragments for the currently observed action, so as to then perform Noisy-OR combination for the action node (Looking). (Box B) Selec-
tion of Bayesian Network fragments for the intentions. (Box C) Constructing a three-layer Bayesian Network, upon which intention recognition
is performed.
AUTHOR COPY
240 T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making
Furthermore, from the design point of view, it is easier
and usually much cheaper to construct the small frag-
ments (and then combine them) than to construct the
whole BN [51,61].
2. Decision making with intention recognition
Given the crucial role that intentions play in a diver-
sity of decision making processes [8,57,75,83,92], one
would expect intentions to occupy a substantial place
in any theory of action. But surprisingly, in one of the
most influential theory of action – the rational choice
theory [5,77] – including the theory of decision mak-
ing – explicit reference is made to actions, strategies,
information, outcomes and preferences, but not to in-
tentions.
This is not to say that no attention has been paid to
the relationship between rational choice and intentions.
Quite the contrary, a rich philosophical and Artificial
Intelligence (AI) literature has developed on the re-
lation between rationality and intentions [8,19,85,89].
Some philosophers, e.g. in [8,10,75], have been con-
cerned with the role that intention plays in directing
rational decision making and guiding future actions.
In addition, many agent researchers have recognized
the importance of intentions in developing useful agent
theories, architectures, and languages, such as Rao and
Georgeff with their BDI model [73], which has led to
the commercialization of several high-level agent lan-
guages, e.g. in [14,93,94]. However, to the best of our
knowledge, there has been no real attempt to model
and implement the role of intentions in decision mak-
ing, within a rational choice framework. Intentions of
other relevant agents are always assumed to be given
as the input of a decision making process; no system
that integrates a real intention recognition system into
a decision making system has been implemented so far.
In our work, we have set forth a coherent Logic Pro-
gramming (LP) based system for decision making –
which extends the existing work on Evolution Prospec-
tion for decision making [63,64] – but taking into con-
sideration now the intentions of other agents. Obvi-
ously, when being immersed in a multi-agent system,
knowing the intentions of other agents can benefit the
agent in a number of ways. It enables the recogniz-
ing agents to predict what other agents will do next or
might have done before. Hence, they can plan in ad-
vance and take the best advantage from the prediction,
or act to take a remedial action. In addition, an impor-
tant role of recognizing intentions is to enable coordi-
nation of your own actions and in collaborating with
others [8,49].
The Evolution Prospection (EP) system is an im-
plemented LP-based system for decision making [63,
64,67]. It is implemented on top of XSB Prolog [96],
a full account of which can be found in [64]. An EP
agent can prospectively look ahead a number of steps
into the future to choose the best course of evolution
that satisfies a goal. This is achieved by designing and
implementing several kinds of prior and post prefer-
ences, and several useful environment-triggering con-
structs for decision making. In order to take into ac-
count the intentions of other agents in decision mak-
ing processes, we integrated into EP a previously and
separately implemented, but also LP-based, intention
recognition system [35,65,67].
The obtained integrated system can perform inten-
tion-based decision making [36,37]. It takes into ac-
count recognized intentions of other agents within dif-
ferent constructs of the decision making system, no-
tably intention-triggering goals (e.g. upon recognizing
intentions of a friend, the goal of helping him/her is
provoked, while upon recognizing intentions of an en-
emy, the goal of preventing his/her achievement of the
recognized intention is triggered) and different kinds of
intention-triggering preferences (e.g. upon recognizing
an intention of a friend, one may prefer an action to
another one if it provides more support to achieve the
intention; in contrast, if it is an enemy, the ones pro-
viding greater support are disfavored).
The system has been applied for providing appro-
priate assistance for elderly people in the Ambient In-
telligence domain [32,33,66] (e.g. upon recognizing
the intention of an elderly person staying alone in his
apartment, the system derives suggestions on how to
achieve his intentions appropriately, taken into account
his profiles and different aspects of the living envi-
ronment); for deriving morally acceptable decisions in
moral dilemmas [37,42] (it is known that a key factor
in legal and moral judgments is actual intention, which
for instance can distinguish murder from manslaughter
[43,97]).
Furthermore, our recent work has paid much atten-
tion to the problem of intention-based decision mak-
ing in large-scale multi-agent systems [38–41]. We
study the performance of agents which are capable of
intention recognition within a population of interact-
ing agents – in order to investigate what is the role
of intention recognition in the evolution of coopera-
tive behaviors [60,84]. Interestingly, we find that such
recognizing agents can learn to cooperate with each
AUTHOR COPY
T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making 241
other in an environment where cheating is favorable.
Therein intention-based decision making is applied in
the course of social dilemmas (e.g. the famous Pris-
oner’s Dilemma [84]). Note that this work significantly
diverges from the existent AI literature on intention
recognition (see below), wherein the study is carried
out at small, local settings: how to efficiently and cor-
rectly recognize particular agents’ intention. As such,
our work has contributed not only to the evolutionary
and philosophical studies of intention recognition [68],
it has suggested an approach to study and evaluate in-
tention recognition methods in large-scale social and
biological complex systems.
Example 2.1 (Intention-triggering goal: Elder care).
Suppose in the previous example the intention recogni-
tion system predicts that the elder intends to find some-
thing to drink. The assistant system provides a sugges-
tion on what kind of drink the elder should take. Tea or
coffee? The following (simplified) EP program targets
that.
on_observe(suggest)<- has_intent(elder,drink).
suggest <- tea. suggest <- coffee.
expect(tea). expect(coffee).
expect_not(coffee) <- blood_high_pres(elder).
coffee <| tea <- sleepy(elder).
The first line reads: if it is recognized that the elder
has the intention of finding something to drink, then
the goal of suggesting an appropriate drink is triggered.
The second line says, tea and coffee are the possi-
ble suggestion options. The third line states, both op-
tions are always expectable. An exception is that coffee
is prohibited when the elder has high blood pressure
(fourth line). Furthermore, coffee is preferred to tea –
when both options are still available after considering
all other constraints – provided the elder is sleepy (fifth
line).
Example 2.2 (Intention-triggering preferences). Be-
ing thirsty, I consider making tea or coffee. I realize
that my roommate, John, also wants to have a drink.
To be friendly, I want to take into account his intention
when making my choice. This scenario is represented
with the following preference rules in EP.
tea <| coffee <- has_intent(john,tea).
coffee <| tea <- has_intent(john,coffee).
For further details, and more complete and extended
examples, see [37,66].
3. Related work on intention recognition
Work on intention and plan recognition has been
paid much attention for more than thirty years, and a
large number of methods have been applied. They can
be roughly categorized into two main groups: Consis-
tency and Probabilistic approaches [2,26,79,86].
Consistency approaches face the problem by deter-
mining which intention is consistent with the observed
actions, i.e. whether the observed actions match with at
least a plan achieving the intention. The earliest work
on plan recognition belongs to this group [45,50,80,81,
91]. More recent work can be found in a rather com-
prehensive survey by [79]. The problem with the con-
sistency approaches is that they cannot handle well the
case where the current observed actions enable more
than one intention – they cannot directly select be-
tween those intentions.
Probabilistic approaches, on the other hand, are
mainly based on Bayesian Network and (Hidden)
Markov models [1,3,13,17,22,26,47,65,67,71,82,88].
An advantage of the probabilistic approaches is that
they can directly address the above issue of the con-
sistency approaches – by finding the most probable in-
tentions given the set of current observations, on the
basis of accumulated statistical evidence, or simply on
the basis of subjective beliefs encoded in a Bayesian
Network or Markov model.
Bayesian approaches have been one of the most suc-
cessful models applied to the intention/plan recogni-
tion problem [17,24,26,28,71]. The first model was
built by [16,17]. Depending on the structure of plan li-
braries, a knowledge-based model construction is em-
ployed to build BNs from the library – which is then
used to infer the posterior probability of explanations
(for the set of observed actions). This approach, mostly
advanced by [28] and especially in the more recent
work [26],1addresses a number of issues in inten-
tion/plan recognition, e.g., when the observed agent
follows multiple intentions or interleaved plans simul-
taneously; fails to observe actions; addresses partly or-
dered plans. However, there are some important as-
pects not yet explored therein, partially for the sake
of computational efficiency. First, prior probabilities
of intentions are assumed to be fixed. This assump-
tion is not always reasonable because those prior prob-
abilities should in general depend on the situation at
hand [8,9,11,71], and can justifiably be captured by
1Note that this work is based on Bayesian inference, though they
do not build Bayesian Networks as in [16,17].
AUTHOR COPY
242 T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making
causes/reasons of the intentions, as in our method
[38,44,67,88]. Indeed, [26] also highlighted the need
to account for contextual information or state of the
world as a potential extension to their plan recog-
nizer. In [71], a similar context-dependent Bayesian
approach is used, though the model therein is not in-
cremental. The authors demonstrated that taking into
account contextual information is important to appro-
priately recognize drivers’ intention in the traffic mon-
itoring domain [71].
Second, intentions are assumed to be independent of
each other. This is not generally the case since the in-
tentions may support or exclude one another, leading to
the need to reconfigure the model. Hence, those works
might not appropriately address multiple intentions
recognition. Pynadath and Wellman [71] proposed to
combine, in their BN model for plan recognition, the
mutually exclusive plan nodes into a single variable.
As a step further, we formally define how that can be
done appropriately, so as to guarantee consistency in
the obtained BN. This latter assumption must always,
explicitly or implicitly, be made by the approaches
based on (Hidden) Markov models, e.g. [3,12], or sta-
tistical corpus-based machine learning [6,7]. Gener-
ally, in those approaches, a separate model is built for
each intention; thus no relations amongst the intentions
are expressed or can be expressed. These works were
restricted to the single intention case. The method de-
veloped in our work attempts to tackle the multiple
case more appropriately. We plan on further experi-
mentation for evaluating it. In any case, note that al-
though there were some previous attempts, e.g., in [26]
and [71], no experiments have been carried out.
Different from most above mentioned works, our
model is context-dependent, which is achieved by in-
cluding in it causes/reasons of intentions. This way,
our model can appropriately deal with the abandon-
ment/changes of intentions – when the causes/reasons
do not support or force the intending agent to hold
those intentions anymore – in an integrated manner. In
contrast, in [25], the authors build a separate model to
recognize when the observed agent abandons its cur-
rent intention, which may then trigger revision of the
intention recognition model. To the best of our knowl-
edge, this is the only work addressing the abandonment
issue. However, the system presented therein is only
evaluated with a rather small benchmark (with three
intentions), and only for the accuracy of the abandon-
ment recognition itself. The benefit from having this
additional intention abandonment recognition module
for enhancing intention/plan recognition performance
has not been studied, as the authors themselves men-
tion in their recent study [26]. We also address this is-
sueinourwork.
In line with our recent work is our own prior work,
a context-dependent incremental intention recognition
model [31]. But there we only deal with the single in-
tention case, and it has exponential complexity. Our
latest model is more general and efficient, being able
to deal with the multiple intention case, and has linear
complexity for the single intention case [34].
The method is performed by dynamically construct-
ing a three-layer BN model for intention recognition
(IRBN), from a prior knowledge base consisting of
readily maintained and constructed fragments of BN.
Their simple structures allow easy maintenance by do-
main experts and also facilitate automatic construc-
tion from available plan corpora. The three-layer IRBN
follows a causal intentional structure [44,88], from
causes/reasons (of intentions) to intentions (as causes
of actions), then to (actually observed) actions. This
causal structure enabled us to appropriately capture
different aspects of context-dependent intention recog-
nition, from context-dependent selection of intentions
for model construction, to sensitizing prior probabili-
ties of the intentions already in the model.
In our work we have provided several operators for
constructing as well as simplifying the BN model dy-
namically, highlighting the importance of taking into
account contextual information – an aspect usually
omitted in previous approaches [26,71]. In addition,
we have shown experimentally that contextual infor-
mation is crucial for recognition accuracy when the
observed agents might change or abandon their initial
goals/intentions. It is apparently an unavoidable aspect
of real agents, clearly pointed out in [25,26]. Taking
into account such information enables to account for
the reasons why the agents change or abandon their
goals/intentions.
We have also attempted to tackle the problem where
an observed agent may follow multiple intentions
simultaneously, in a more appropriate manner. We
have formally described how to represent relationships
amongst intentions in the Bayesian Network for inten-
tion recognition, particularly in order to maintain its
consistency when one needs to combine mutually ex-
clusive intentions [71]. This aspect is indispensable in
multiple intentions recognition, but mostly omitted in
previous work. However, the scalability of our method
remains to be seen. For its evaluation, we still need to
gather an appropriate plan corpus allowing for the pos-
sibility that users pursue multiple intentions simultane-
ously.
AUTHOR COPY
T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making 243
A limitation of the current formalization in the mul-
tiple intentions recognition case is that we need to as-
sume that the intentions to be combined are perfectly
mutually exclusive. This assumption can be relaxed by
utilizing a latent variable for any subset of perfectly
mutually exclusive intention nodes. The latent vari-
able figures in the BN, either as a child or parent of
the nodes, whichever works better for inference. We
are exploring this direction to provide a more general
method for representing relationships amongst inten-
tion nodes.
Another limitation of our current method is that it
did not explicitly take into account temporal evolu-
tion of domain variables in the BN. It is usually done
using Dynamic Bayesian Networks (DBNs) [1,22,71],
where the state of each variable is represented by a
series of nodes. In our method, the time evolution, to
some degree, is implemented by means of updating the
IRBN from time to time. For example, the states of the
cause/reason nodes are updated using an external logic
program which represents the evolving world. In this
way not only can one significantly reduce the size of
the BN, and thereby its inference complexity (as LP in-
ference is significantly less expensive), but the declara-
tive representation and reasoning of LP techniques [4]
could be important when the states of a node cannot
be easily represented in an explicit manner. However,
in domains where the states of the nodes might con-
stantly change, the explicit time series representation
of DBNs is apparently necessary. We envisage to bring
in, to some degree, the time series representation of
DBNs [1,20,22], to improve our method.
One future direction area we aim at is the real de-
ployment of our intention recognition method to tackle
different real application domains, e.g., Ambient Intel-
ligence [33,78] and Elder Care [66], where intention
recognition has been of increasing importance [33,78,
79].
4. Related work on intention recognition for
decision making
Many issues concerning intentions have been widely
discussed in the literature of agent research. Some
philosophers, e.g., Bratman [8,10] have been con-
cerned with the role that intention plays in directing
rational decision making and guiding future actions.
Many agent researchers have recognized the impor-
tance of intentions in developing useful agent theories,
architectures, and languages, such as Rao and Georgeff
with their BDI model [73], which, as we said above,
has led to the commercialization of several high-level
agent languages (e.g., see [14,94]).
However, to the best of our knowledge, there has
been no real attempt to model and implement the role
of intentions in decision making, within a rational
choice framework. Intentions of other relevant agents
are always assumed to be given as the input of a deci-
sion making process; no system that integrates a real
intention recognition system into a decision making
system has been implemented so far.
The ongoing work of [33,66,67] also attempts to
combine the two systems, Evolution Prospection and
Intention Recognition, but in a completely different
manner. Their we use an intention recognition system
to recognize the goal of the observed agent (e.g., an el-
der [66]), which the evolution prospection system then
uses to derive appropriate courses of actions to help
achieve.
Our recently published approach is more general
and genuinely integrated: the intention recognition sys-
tem is employed also to evaluate other different kinds
of information being utilized, within an EP program
[36,37]. It summarizes the existent work on Evolution
Prospection and Intention Recognition and shows a co-
herent combination of them for decision making. The
Evolution Prospection system has been proven to be a
useful one for decision making, and now it has been
empowered to take into account intentions of other
agents – an important aspect that had not been explored
so far. The fact that both systems are LP-based has en-
abled their easy integration.
Acknowledgements
TAH acknowledges the support from FCT-Portugal
(grant SFRH/BD/62373/2009) and the FWO Vlaan-
deren, Belgium (Postdoctoral fellowship).
References
[1] D.W. Albrecht, I. Zukerman and A.E. Nicholson, Bayesian
models for keyhole plan recognition in an adventure game,
User Modeling and User-Adapted Interaction 8(1998), 5–47.
[2] M.G. Armentano and A. Amandi, Plan recognition for inter-
face agents, Artif. Intell. Rev. 28 (2007), 131–162.
[3] M.G. Armentano and A. Amandi, Goal recognition with
variable-order Markov models, in: Proceedings of the 21st In-
ternational Joint Conference on Artificial Intelligence, 2009,
pp. 1635–1640.
AUTHOR COPY
244 T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making
[4] C. Baral, Knowledge Representation, Reasoning, and Declar-
ative Problem Solving, Cambridge Univ. Press, New York, NY,
2003.
[5] K. Binmore, Rational Decisions, Gorman Lectures in Eco-
nomics, Princeton Univ. Press, Princeton, NJ, 2009.
[6] N. Blaylock and J. Allen, Corpus-based, statistical goal
recognition, in: Proceedings of the 18th International Joint
Conference on Artificial Intelligence (IJCAI’03), 2003,
pp. 1303–1308.
[7] N. Blaylock and J. Allen, Statistical goal parameter recogni-
tion, in: Proceedings of the 14th International Conference on
Automated Planning and Scheduling (ICAPS’04), S. Zilber-
stein, J. Koehler and S. Koenig, eds, AAAI, 2004, pp. 297–304.
[8] M.E. Bratman, Intention, Plans, and Practical Reason,The
David Hume Series, CSLI, 1987.
[9] M.E. Bratman, Planning and the stability of intention, Minds
and Machines 2(1) (1992), 1–16.
[10] M.E. Bratman, Faces of Intention: Selected Essays on Intention
and Agency, Cambridge Univ. Press, New York, NY, 1999.
[11] S.M. Brown, A Decision Theoretic Approach for Interface
Agent Development, PhD thesis, Faculty of the Graduate
School of Engineering of the Air Force Institute of Technology,
Air University, USA, 1998.
[12] H. Bui, A general model for online probabilistic plan recogni-
tion, in: Proceedings of the 8th International Joint Conference
on Artificial Intelligence (IJCAI’03), G. Gottlob and T. Walsh,
eds, 2003, pp. 1309–1318.
[13] H. Bui, S. Venkatesh and G. West, Policy recognition in the ab-
stract hidden Markov model, Journal of Artificial Intelligence
Research 17 (2002), 451–499.
[14] B. Burmeister, M. Arnold, F. Copaciu and G. Rimassa, BDI-
agents for agile goal-oriented business processes, in: Proceed-
ings of the 7th International Joint Conference on Autonomous
Agents and Multiagent Systems: Industrial Track, AAMAS’08,
2008, pp. 37–44.
[15] E. Charniak and R.P. Goldman, Plan recognition in stories
and in life. In Proceedings of the Fifth Annual Conference on
Uncertainty in Artificial Intelligence, UAI’89, North-Holland,
Amsterdam, The Netherlands, 1990, pp. 343–352.
[16] E. Charniak and R.P. Goldman, A probabilistic model of plan
recognition, in: Proceedings of the Ninth National Confer-
ence on Artificial Intelligence (AAAI’1991), AAAI’91, Vol. 1,
AAAI Press, 1991, pp. 160–165.
[17] E. Charniak and R.P. Goldman, A Bayesian model of plan
recognition, Artificial Intelligence 64(1) (1993), 53–79.
[18] D.L. Cheney and R.M. Seyfarth, Baboon Metaphysics: The
Evolution of a Social Mind, University of Chicago Press,
Chicago, IL, 2007.
[19] P.R. Cohen and H.J. Levesque, Intention is choice with com-
mitment, Artificial Intelligence 42(2,3) (1990), 213–261.
[20] T. Dean and K. Kanazawa, Probabilistic Temporal reasoning,
in: Proceedings of the Seventh National Conference on Artifi-
cial Intelligence (AAAI-88), AAAI, 1992, pp. 524–528.
[21] F.J. Diez, Parameter adjustment in Bayes networks, the gener-
alized noisy or-gate, in: Proceedings of Conference on Uncer-
tainty in Artificial Intelligence (UAI’93), Morgan Kaufmann,
San Francisco, CA, 1993, pp. 99–105.
[22] J. Forbes, T. Huang, K. Kanazawa and S. Russell, The batmo-
bile: towards a Bayesian automated taxi, in: Proceedings of the
14th International Joint Conference on Artificial Intelligence,
Vol. 2, 1995, pp. 1878–1885.
[23] C.W. Geib, Problems with intent recognition for elder care,
in: Proceedings of AAAI Workshop Automation as Caregiver,
2002, pp. 13–17.
[24] C.W. Geib, Assessing the complexity of plan recognition, in:
Proceedings of the Conference of the American Association of
Artificial Intelligence (AAAI’2004), 2004, pp. 507–512.
[25] C.W. Geib and R.P. Goldman, Recognizing plan/goal aban-
donment, in: Proceedings of the 18th International Joint
Conference on Artificial Intelligence (IJCAI’03), 2003,
pp. 1515–1517.
[26] C.W. Geib and R.P. Goldman, A probabilistic plan recognition
algorithm based on plan tree grammars, Artificial Intelligence
173(2009) (2009), 1101–1132.
[27] V. Gogate and R. Dechter, SampleSearch: Importance sam-
pling in presence of determinism, Artificial Intelligence 175(2)
(2011), 694–729.
[28] R.P. Goldman, C.W. Geib and C. Miller, A new model of plan
recognition, in: Proceedings of Conference on Uncertainty in
Artificial Intelligence (UAI’99), 1999, pp. 245–254.
[29] K. Haigh, L. Kiff, J. Myers, V. Guralnik, C. Geib, J. Phelps and
T. Wagner, The Independent LifeStyle Assistant (I.L.S.A.): AI
lessons learned, in: Proceedings of Conference on Innovative
Applications of Artificial Intelligence, 2004, pp. 852–857.
[30] T.A. Han, Intention Recognition, Commitments and Their
Roles in the Evolution of Cooperation, From Artificial In-
telligence Techniques to Evolutionary Game Theory Models,
Vol. 9, Springer, Heidelberg, 2013.
[31] T.A. Han and L.M. Pereira, Anytime intention recognition via
incremental Bayesian network reconstruction, in: AAAI 2010
Fall Symposium on Proactive Assistant Agents (PAA 2010),
AAAI, 2010, pp. 20–25.
[32] T.A. Han and L.M. Pereira, Collective intention recognition
and elder care, in: AAAI 2010 Fall Symposium on Proactive
Assistant Agents (PAA 2010), AAAI, 2010, pp. 26–31.
[33] T.A. Han and L.M. Pereira, Proactive intention recognition for
home ambient intelligence, in: Proceedings of 5th Workshop
on Artificial Intelligence Techniques for Ambient Intelligence
(AITAmI’10), Ambient Intelligence and Smart Environments,
Vol. 8, IOS Press, 2012, pp. 91–100.
[34] T.A. Han and L.M. Pereira, Context-dependent incremental de-
cision making scrutinizing intentions of others via Bayesian
network model construction, 2012, submitted.
[35] T.A. Han and L.M. Pereira, Context-dependent incremental in-
tention recognition through Bayesian network model construc-
tion, in Proceedings of the Eighth UAI Bayesian Modeling Ap-
plications Workshop (UAI-AW 2011), A. Nicholson, ed., CEUR
Workshop Proceedings, Vol. 818, 2011, pp. 50–58.
[36] T.A. Han and L.M. Pereira, Intention-based decision mak-
ing with evolution prospection, in: Proceedings of 15th Por-
tuguese Conference on Artificial Intelligence (EPIA’2011),
L. Antunes and H. Pinto, eds, LNAI, Vol. 7026, Springer, 2011,
pp. 254–267.
[37] T.A. Han and L.M. Pereira, Intention-based decision making
and its applications, in: Human Behavior Recognition Tech-
nologies: Intelligent Applications for Monitoring and Secu-
rity, H. Guesgen and S. Marsland, eds, IGI Global, 2013,
pp. 174–211 (Chapter 9).
[38] T.A. Han, L.M. Pereira and F.C. Santos, Intention recognition
promotes the emergence of cooperation, Adaptive Behavior
19(3) (2011), 264–279.
AUTHOR COPY
T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making 245
[39] T.A. Han, L.M. Pereira and F.C. Santos, The role of intention
recognition in the evolution of cooperative behavior, in: Pro-
ceedings of the 22nd International Joint Conference on Arti-
ficial Intelligence (IJCAI’2011), T. Walsh, ed., AAAI, 2011,
pp. 1684–1689.
[40] T.A. Han, L.M. Pereira and F.C. Santos, Corpus-based inten-
tion recognition in cooperation dilemmas, Artificial Life Jour-
nal 18(4) (2012), 365–383.
[41] T.A. Han, L.M. Pereira and F.C. Santos, Intention recognition,
commitment, and the evolution of cooperation, in: Proceedings
of IEEE Congress on Evolutionary Computation, IEEE Press,
2012, pp. 1–8.
[42] T.A. Han, A. Saptawijaya and L.M. Pereira, Moral reasoning
under uncertainty, in: Proceedings of the 18th International
Conference on Logic for Programming, Artificial Intelligence
and Reasoning (LPAR-18), LNAI, Vol. 7180, Springer, 2012,
pp. 212–227.
[43] M.D. Hauser, Moral Minds, How Nature Designed Our Uni-
versal Sense of Right and Wrong, Little Brown, London, UK,
2007.
[44] C. Heinze, Modeling Intention Recognition for Intelligent
Agent Systems, PhD thesis, The University of Melbourne, Aus-
tralia, 2003.
[45] J. Hong, Goal recognition through goal graph analysis, Journal
of Artificial Intelligence Research 15 (2001), 1–30.
[46] E. Horvitz, J. Breese, D. Heckerman, D. Hovel and K. Rom-
melse, The lumière project: Bayesian user modeling for infer-
ring the goals and needs of software users, in: Proceedings of
the 14th Conference on Uncertainty in Artificial Intelligence,
G.F. Cooper and S. Moral, eds, Morgan Kaufmann, San Fran-
cisco, CA, 1998, pp. 256–265.
[47] M. Huber and R. Simpson, Recognizing the plans of screen
reader users, in: Workshop on Modeling Other Agents from Ob-
servations (MOO2004), 2004.
[48] M. Iacoboni, I. Molnar-Szakacs, V. Gallese, G. Buccino,
J.C. Mazziotta and G. Rizzolatti, PLoS Biol. 3(3) (2005), e79.
[49] G.A. Kaminka, D.V.P.M. Tambe, D.V. Pynadath and
M. Tambe, Monitoring teams by overhearing: A multi-agent
plan-recognition approach, Journal of Artificial Intelligence
Research 17 (2002), 83–135.
[50] H. Kautz and J. Allen, Generalized plan recognition, in: Pro-
ceedings of the Conference of the American Association of Ar-
tificial Intelligence (AAAI’1986), AAAI, 1986, pp. 32–38.
[51] K. Laskey, MEBN: A language for first-order Bayesian knowl-
edge bases, Artificial Intelligence 172(2,3) (2008), 140–178.
[52] K. Laskey and S. Mahoney, Network fragments: Representing
knowledge for constructing probabilistic models, in: Proceed-
ings of Conference on Uncertainty in Artificial Intelligence
(UAI’97), Morgan Kaufmann, San Francisco, CA, 1997.
[53] N. Lesh, Scalable and adaptive goal recognition, PhD thesis,
University of Washington, 1998.
[54] O. Madani, H. Bui and E. Yeh, Efficient online learning and
prediction of users’ desktop actions, in: Proceedings of the
21nd International Joint Conference on Artificial Intelligence
(IJCAI’2009), 2009, pp. 1457–1462.
[55] S.M. Mahoney and K.B. Laskey, Network engineering for
complex belief networks, in: Proceedings of Conference on
Uncertainty in Artificial Intelligence (UAI’96),MorganKauf-
mann, San Francisco, CA, 1996, pp. 389–396.
[56] W. Mao and J. Gratch, A utility-based approach to intention
recognition, in: AAMAS 2004 Workshop on Agent Tracking:
Modeling Other Agents from Observations,2004.
[57] A.N. Meltzoff, Imitation and other minds: the “like me” hy-
pothesis, in: Perspectives on Imitation: From Neuroscience to
Social Science. Imitation, Human Development, and Culture,
MIT Press, Cambridge, MA, 2005, pp. 55–77.
[58] K. Nakahara and Y. Miyashita, Understanding intentions:
Through the looking glass, Science 308(5722) (2005),
644–645.
[59] S. Natarajan, P. Tadepalli, T.G. Dietterich and A. Fern, Learn-
ing first-order probabilistic models with combining rules, An-
nals of Mathematics and Artificial Intelligence 54 (2008),
223–256.
[60] M.A. Nowak, Five rules for the evolution of cooperation, Sci-
ence 314(5805) (2006), 1560.
[61] J. Pearl, Probabilistic Reasoning in Intelligent Systems: Net-
works of Plausible Inference, Morgan Kaufmann, San Fran-
cisco, CA, 1988.
[62] J. Pearl, Causality: Models, Reasoning, and Inference,Cam-
bridge Univ. Press, New York, NY, 2000.
[63] L.M. Pereira and T.A. Han, Evolution prospection, in: Pro-
ceedings of International Symposium on Intelligent Decision
Technologies (KES-IDT’09), Springer Studies in Computa-
tional Intelligence, Vol. 199, 2009, pp. 51–63.
[64] L.M. Pereira and T.A. Han, Evolution prospection in deci-
sion making, Intelligent Decision Technologies 3(3) (2009),
157–171.
[65] L.M. Pereira and T.A. Han, Intention recognition via causal
Bayes networks plus plan generation, in: Progress in Arti-
ficial Intelligence, Proceedings of 14th Portuguese Interna-
tional Conference on Artificial Intelligence (EPIA’09),LNAI,
Vol. 5816, Springer, 2009, pp. 138–149.
[66] L.M. Pereira and T.A. Han, Elder care via intention recognition
and evolution prospection, in: Select Extended Papers from the
18th International Conference on Applications of Declarative
Programming and Knowledge Management (INAP’09),LNAI,
Vol. 6547, Springer, 2011, pp. 170–187.
[67] L.M. Pereira and T.A. Han, Intention recognition with evolu-
tion prospection and causal Bayesian networks, in: Computa-
tional Intelligence for Engineering Systems 3: Emergent Ap-
plications, Springer, 2011, pp. 1–33.
[68] L.M. Pereira, T.A. Han and F.C. Santos, Complex systems of
mindful entities – on intention recognition and commitment,
in: Model-Based Reasoning in Science and Technology: The-
oretical and Cognitive Issues, L. Magnani, ed., Springer, in
press.
[69] A. Pfeffer, D. Koller, B. Milch and K.T. Takusagawa, Pook:
A system for probabilistic object-oriented knowledge repre-
sentation, in: Proceedings of Conference on Uncertainty in Ar-
tificial Intelligence (UAI’93), 1999, pp. 541–550.
[70] D. Premack and G. Woodruff, Does the chimpanzee have a
theory of mind?, Behavioral and Brain Sciences 1(4) (1978),
515–526.
[71] D.V. Pynadath and M.P. Wellman, Accounting for context in
plan recognition, with application to traffic monitoring, in:
Proceedings of Conference on Uncertainty in Artificial Intelli-
gence (UAI’95), 1995, pp. 472–481.
AUTHOR COPY
246 T.A. Han and L.M. Pereira / State-of-the-art of intention recognition and its use in decision making
[72] M. Ramírez and H. Geffner, Probabilistic plan recognition us-
ing off-the-shelf classical planners, in: Proceedings of the Con-
ference of the American Association of Artificial Intelligence
(AAAI’2010), 2010, pp. 1121–1126.
[73] A.S. Rao and M.P. Georgeff, BDI Agents: from theory to prac-
tice, in: Proceeding of First International Conference on Mul-
tiagent Systems, 1995, pp. 312–319.
[74] G. Rizzolatti and L. Craighero, The mirror-neuron system, An-
nual Review of Neuroscience 27 (2004), 169–192.
[75] O. Roy, Thinking before acting: Intentions, logic, rational
choice, PhD thesis, ILLC Dissertation Series DS-2008-03,
Amsterdam, 2009.
[76] P. Roy, B. Bouchard, A. Bouzouane and S. Giroux, A hybrid
plan recognition model for Alzheimer’s patients: interleaved-
erroneous dilemma, in: Proceedings of IEEE/WIC/ACM Inter-
national Conference on Intelligent Agent Technology, 2007,
pp. 131–137.
[77] S.J. Russell and P. Norvig, Artificial Intelligence: A Modern
Approach, Pearson Education, 2003.
[78] F. Sadri, Ambient intelligence: A survey, ACM Computing Sur-
vey 43 (2011), 36:1–36:66.
[79] F. Sadri, Logic-based approaches to intention recognition, in:
Handbook of Research on Ambient Intelligence: Trends and
Perspectives, 2011, pp. 346–375.
[80] F. Sadri, Intention recognition with event calculus graphs
and weight of evidence, in: Proceedings 4th International
Workshop on Human Aspects in Ambient Intelligence, 2010,
pp. 470–475.
[81] C. Schmidt, N. Sridharan and J. Goodson, The plan recogni-
tion problem: An intersection of psychology and artificial in-
telligence, Artificial Intelligence 11 (1978), 45–83.
[82] O.C. Schrempf, D. Albrecht and U.D. Hanebeck, Tractable
probabilistic models for intention recognition based on expert
knowledge, in: Proceedings of International Conference on In-
telligent Robots and Systems, IEEE, 2007, pp. 1429–1434.
[83] J.R. Searle, Making the Social World: The Structure of Human
Civilization, Oxford Univ. Press, Oxford, UK, 2010.
[84] K. Sigmund, The Calculus of Selfishness, Princeton Univ.
Press, Princeton, NJ, 2010.
[85] M.P. Singh, Intentions, commitments and rationality, in: 13th
Annual Conference of the Cognitive Science Society, 1991,
pp. 493–498.
[86] P. Singla and R.J. Mooney, Abductive Markov logic for plan
recognition, in: Proceedings of the Conference of the Ameri-
can Association of Artificial Intelligence (AAAI’2011), 2011,
pp. 1069–1075.
[87] S. Srinivas, A generalization of the noisy-or model, in: Pro-
ceedings of Conference on Uncertainty in Artificial Intelli-
gence (UAI’93), Morgan Kaufmann, San Francisco, CA, 1993,
pp. 208–218.
[88] K.A. Tahboub, Intelligent human-machine interaction based
on dynamic Bayesian networks probabilistic intention recog-
nition, J. Intell. Robotics Syst. 45 (2006), 31–52.
[89] M. van Hees and O. Roy, Intentions and plans in decision and
game theory, in: Reasons and Intentions, Ashgate Publishers,
2008, pp. 207–226.
[90] A. Whiten, Natural Theories of Mind: Evolution, Develop-
ment, and Simulation of Everyday Mindreading, Blackwell,
Oxford, UK, 1991.
[91] R. Wilensky, Planning and Understanding,Addison-Wesley,
Reading, MA, 1983.
[92] A.L. Woodward, J.A. Sommerville, S. Gerson, A.M. Hender-
son and J. Buresh, The emergence of intention attribution in in-
fancy, in: The Psychology of Learning and Motivation, Vol. 51,
B.H. Ross, ed., Academic Press, 2009, pp. 187–222.
[93] M. Wooldridge, Reasoning about Rational Agents, MIT Press,
Cambridge, MA, 2000.
[94] M. Wooldridge, Reasoning about rational agents, The Journal
of Artificial Societies and Social Simulation 5(1) (2002).
[95] Y.P. Xiang and K.L. Poh, Constructing Bayesian network in
a changing world, in: Proceedings of AAAI Spring Sympo-
sium on Challenges to Decision Support in a Changing World,
AAAI, 2005, pp. 108–113.
[96] XSB, The XSB System Version 3.2, Libraries, interfaces and
packages, Vol. 2, 2009.
[97] L. Young and R. Saxe, When ignorance is no excuse: Differ-
ent roles for intent across moral domains, Cognition 120(2)
(2011), 202–214.
