Controlling Narrative Generation with Planning Trajectories: The Role of Constraints

Abstract

AI planning has featured in a number of Interactive Storytelling prototypes: since narratives can be naturally modelled as
a sequence of actions it is possible to exploit state of the art planners in the task of narrative generation. However the
characteristics of a “good” plan, such as optimality, aren’t necessarily the same as those of a “good” narrative, where errors
and convoluted sequences may offer more reader interest, so some narrative structuring is required. We have looked at injecting
narrative control into plan generation through the use of PDDL3.0 state trajectory constraints which enable us to express
narrative control information within the planning representation. As part of this we have developed an approach to planning
with trajectory constraints. The approach decomposes the problem into a set of smaller subproblems using the temporal orderings
described by the constraints and then solves them incrementally. In this paper we outline our method and present results that
illustrate the potential of the approach.

Full text

Controlling Narrative Generation with Planning
Trajectories: the Role of Constraints
Julie Porteous and Marc Cavazza
School of Computing
University of Teesside
{j.porteous}{m.o.cavazza}@tees.ac.uk
Abstract. AI planning has featured in a number of Interactive Story-
telling prototypes: since narratives can be naturally modelled as a se-
quence of actions it has been possible to exploit state of the art plan-
ners in the task of narrative generation. However the characteristics of
a “good” plan, such as optimality, aren’t necessarily the same as those
of a “good” narrative, where errors and convoluted sequences may offer
more reader interest, so some narrative structuring is required. In our
work we have looked at injecting narrative control into plan generation
through the use of PDDL3.0 state trajectory constraints which enable
us to express narrative control information within the planning repre-
sentation. As part of this we have developed an approach to planning
with such trajectory constraints. The approach decomposes the problem
into a set of smaller subproblems using the temporal orderings described
by the constraints and then solves these subproblems incrementally. In
this paper we outline our method and present results that illustrate the
potential of the approach.
1 Introduction
A key component of an Interactive Storytelling (IS) system is a narrative gen-
erator. Since narratives can be naturally modelled as a sequence of actions, AI
planning has emerged as the technology of choice in this field and a range of
planning approaches have been successfully applied to the task of narrative gen-
eration. For example, partial order planning in the tradition of UCPOP [1] has
been used with adaptations to intentions [2] and emotion [3]; state based plan-
ning in the style of HSP [4] was used to generate plans in an authoring tool [5];
and HTN planning [6] was adapted to handle anytime user intervention [7].
As these systems demonstrate, it is clearly possible to generate narratives us-
ing planning technology but they are not without their limitations. For instance,
both partial order (plan space) planning and state space planning approaches
suffer from an IS perspective, because they were developed with the objective of
building a planner that would find the shortest plan to a goal (or a close approxi-
mation to it). This is appropriate in the sorts of domains for which planners were
first developed, such as robot stacking and logistics, but has little meaning in

2
the context of IS, where errors and convoluted sequences may be desirable char-
acteristics. On the other hand, HTN planning does provide a means to include
information about desirable ways for the narrative to be developed, in the form
of decompositions of methods in the task network. However, because the control
information is embedded in the task network it is not possible to manipulate
the information independently and since this is a feature that would be useful in
an IS environment, for example to respond to user interaction, it suggests some
other approach to narrative structuring is required.
A review of the IS literature reveals a number of previous attempts at nar-
rative structuring. The approach taken by some has been to use HTN planning
which, as previously mentioned, permits the encoding of authors intentions of
possible ways to develop the narrative, in the form of the decomposition methods
themselves [7, 8]. Others have proposed the use of search based drama manage-
ment: Weyhrauch [9] suggested the encoding of aesthetics in an evaluation func-
tion to be used by the drama manager; Magerko et al [10] exploited a meta-level
controller in the form of a drama manager; Mateas and Stern’s Fa¸cade drama
[11] featured a beat based drama manager; and Riedl and Stern [12] used an au-
tomated story director to blend user agency with narrative plot control. Other
researchers have looked to adapt the plan generation algorithm itself to struc-
ture the narrative along aesthetic lines. For example Riedl and Young [2], used
simulated intention reasoning to guide the planner to promote story coherence
and temporal consistency whilst Cheong and Young [8] demonstrated how the
narrative structure can be post-processed to help promote suspense.
In our work we have looked at an alternate way to structure the narrative:
by forcing the planner to make certain key events occur in the course of the
narrative and in a given order. Riedl [13] referred to this as “complexifying” the
planning process and extended his intent driven planner [2] to plan with the
inclusion of “author goals”. However, we have taken a rather different approach
based on the observation that the state trajectory constraints provided in the
planning domain representation language PDDL3.0 appear to provide a means
to express the key events that are required in a narrative and also any ordering
between them. Our hypothesis is that these constraints can be used to express
narrative structuring control knowledge and that any narratives generated that
satisfy these constraints will display the desired narrative structure. We were
encouraged by the results of our initial experiments with MIPS-xxl [14], a planner
that can reason about PDDL3.0 constraints, which supported our hypothesis.
However that planner appeared too slow, given IS requirements, and this led us
to develop a novel method for planning with these constraints that could perform
within the time limits imposed by an IS environment. In the paper, we present
an overview of this work and some preliminary results.
The paper is organised as follows. We begin in section 2 with background on
the use of constraints to inject narrative control into story generation. Then in
section 3 we discuss the novel algorithm that we have developed for planning
with trajectory constraints. In section 4 we discuss our results and at the end of
the paper in section 5 we conclude and suggest directions for future work.

3
2 Narrative Structuring using Constraints
To illustrate the use of constraints we’ll use an example from an IS domain
based on the novel Goldfinger by Ian Fleming (with a history both in narrative
formalism [15] and IS [16]). The novel features James Bond, a British secret
service agent who is sent to investigate gold smuggling by the eponymous Auric
Goldfinger. Suppose that we want to generate a narrative with a goal of Bond
defusing a bomb set by Goldfinger. We could do this using a planner but a
problem with the narrative, from an IS perspective, is that it will only contain
actions that are directly relevant in pursuit of the planning goal. Yet the narrative
may be more interesting if, for example, James Bond were placed in a perilous
situation (such as facing death by a laser beam), but such events would only
feature in the narrative if they served some purpose with respect to achieving
the final goal (such as Bond gaining further information). What is required is a
means to specify important events that are to occur in the narrative, as well as
the final narrative goals. Our hypothesis is that the state trajectory constraints
in the planning representation language PDDL3.0 offer a way to do this.
State trajectory constraints are one of the novel language features introduced
in PDDL3.0 [17] to describe benchmark problems for the 5th International Plan-
ning Competition. They are expressed independently to the rest of the PDDL
problem definition and assert conditions that must be met by the entire sequence
of states visited during the execution of a solution. The language provides a num-
ber of modal operators for expressing these constraints. In our experiments we
have used the following:
operator meaning
(sometime-before a b) bmust be made true for the first time before a
(sometime a) predicate amust be true at some stage of the narrative
(at-end a) predicate amust be true at the end of the narrative
The sometime and sometime-before operators allow us to specify events that
must occur in the narrative and this gives a temporal order over these events.
For the Goldfinger example these constraints could include:
(sometime-before (seduced bond jill) (won-cards bond goldfinger card-game))
(sometime-before (seduced bond jill) (got-mission bond))
(sometime (won-golf bond goldfinger golf-game))
which gives a partial temporal order. It is partial because not all the events are
ordered with respect to each other: the event (won-golf bond goldfinger) must
appear at some point in the narrative, but the exact timing is left unspecified.
The at-end constraints are equivalent to planning goals, and they are ordered last
in the temporal order (and the initial state is implicitly ordered to be earliest).
An example at-end constraint for Goldfinger could be:
(at-end (defused-bomb bond fort-knox))
The temporal order defined by these constraints is as shown in figure 1 below.

4
Fig. 1. Partial Temporal Order between constrained predicates.
The subset of PDDL3.0 constraints operators that we have discussed fit our
purposes since they enable us to specify both: narrative events that must occur
in the course of the narrative; and any orders between them. They also result in a
number of benefits over alternate approaches. For example, from a computational
perspective, because the constraints are specified independently from the rest of
the PDDL model they can be independently manipulated, perhaps to respond
to user interaction, without affecting the rest of the model. This is in contrast to
approaches where control information is embedded in the operator pre-conditions
or in the decompositions of an HTN and where making changes is much harder.
Also, from the perspective of authoring, the independent specification of the
constraints may make them easier to express and may help facilitate the testing
of different story variants since the constraints can be independently changed
without changing the rest of the representation.
3 Decomposition Planning with Trajectory Constraints
We saw in the previous section that the constraints define a temporal order over
key narrative facts. This temporal order gives strong hints about events that will
feature earlier in any narrative that satisfies the constraints, and suggested to us
that a narrative could be built incrementally, starting with those events which
are likely to feature earlier on. Based on this observation our planning approach
is: use the orders in the constraints to decompose the problem into a number of
temporally ordered subproblems; solve those subproblems in turn; and “grow”
the narrative forwards from the initial state to the goal state by appending the
solution for the current subproblem to the narrative developed so far.
This approach is based on the method for planning with landmarks of Hoff-
mann et al [18]. They exploit landmarks - facts that must necessarily be made
true en route to solving a final problem goal - to decompose the planning prob-
lem and then use the resulting subproblems to guide a search control algorithm
that is wrapped around a base planner (any planner that accepts basic STRIPS
[19] input). Instead of landmarks, we use the events that have been supplied as
constraints to decompose the problem. Although these facts aren’t landmarks as
defined by Hoffmann et al (they don’t have to be made true to achieve the final
goals) they nevertheless have to be achieved in order to satisfy the constraints.
The first step of the planning process is to construct a constraints tree using
the constraints that are included in the input domain model. The input is a

5
Input:F, I , G, C, O Output:P
1build constraints tree CT
2repeat until CT := {}
3D= leaf nodes of constraints tree
4call planner with: A,S, and disjunctive goal D→P0
5if planner didn’t find a solution P0then fail
6else if planner did find a solution P0:
7P:= P◦P0
S:= result of executing P0in S
8remove from CT all L∈Dwith L∈add(o) for some o∈P0
9call planner with: A,Sand conjunctive goal G→P0
10 if planner didn’t find solution P0then fail
11 else if planner did find a solution P0:
12 P:= P◦P0, output P
Fig. 2. Outline Algorithm: Decomposition Search Control
subset of a PDDL3.0 planning problem consisting of: F, a set of facts that can
be used to describe the world; G, a goal condition such that G⊆F; and C, a
(possibly empty) set of constraints. The output is a constraints tree CT with
nodes Nand edges E. The nodes Nin the tree C T are any facts f∈Fthat:
appear in any of the sometime constraints; or any facts that appear as problem
goals or at-end constraints which can be treated as equivalent. For the example
constraints given in section 3 the set of nodes Nis:
{(defused-bomb bond fort-knox), (won-cards bond goldfinger card-game),
(seduced bond jill), (got-mission bond), (won-golf bond goldfinger)}
Edges (x,y) in the tree are directed from xto yand indicate that xmust be
made true in the plan before y. Continuing with the example constraints given
in section 3 the set of edges Eis:
{((won-cards bond goldfinger card-game), (seduced bond jill)),
((got-mission bond), (seduced bond jill)),
((seduced bond jill), (defused-bomb bond fort-knox)),
((won-golf bond goldfinger), (defused-bomb bond fort-knox))}
An outline of the decomposition planning algorithm is shown in figure 2. The
input is a PDDL3.0 planning problem consisting of < F, I , G, C, O > where: F
is a set of facts that can be used to describe the world; Iand Gare an initial
state of the world and goal condition such that I⊆Fand G⊆F;C, a (possibly
empty) set of constraints; and O, a ground set of operators each with an Add,
Delete, and Precondition list. Upon successful termination of the algorithm, the
output is a plan Pthat satisfies the constraints and achieves all final goals.

6
The first step in the algorithm is the construction of the constraints tree,
CT , which we described above. Then the algorithm loops until the constraints
tree is empty (lines 2-8). Within each loop, a new subproblem is formulated
and passed to the base planner (line 3-4). The goal of this subproblem, D, is
disjunctive and is formed from the leaves of the constraints tree: these goals are
the earliest in the temporal order, which we want the base planner to work on
first. We know nothing about whether these facts can be made true at the same
time or their order with respect to each other, so leave it up to the planner
to decide which goal to work on next. We achieve this by expressing them as
disjunctive goals using the method proposed by Gazen and Knoblock [20] which
allows them to be expressed in standard STRIPS format. Then a separate base
planner is invoked with this disjunctive subproblem (line 4). If the base planner
returned a solution, P0, to this disjunctive subproblem then a new plan is formed
by appending P0onto the current plan so far and a new initial state is formed
by executing the returned plan P0in the current subproblem initial state (line
7). Also, the constraints tree CT is updated so that any facts in leaf nodes of
the tree that are made true by the execution of plan P0are removed (line 8).
Once the constraints tree is empty (or if no constraints are supplied) the search
control proceeds with the original top level conjunctive goal G(line 9). When
the plan for this conjunctive goal subproblem is returned this is appended to the
end of the plan so far to produce the final plan which is then output (line 12).
4 Results
The central hypothesis of our work is that PDDL3.0 constraints can be used for
narrative control. In this section we present the results of a qualitative evaluation
that support this hypothesis, via analysis of a selection of sample narratives.
For the evaluation we developed a decomposition planner, referred to as
DPC, which is an implementation of the algorithm outlined in figure 2. DPC
uses FF-v2.3 [21] as a base planner, although any propositional planner that
accepts STRIPS input could have been used. We used DP C to generate sample
narratives for a number of input narrative variants that featured constraints (all
narrative plans generated by DP C were validated using the PDDL3.0 validator
VAL [22] and shown to satisfy the constraints). In these experiments performance
of DPC was found to be acceptable for IS purposes, with an average subproblem
solution time within 500ms.
For the purposes of the evaluation we developed a PDDL3.0 representation
of the classic spy novel Goldfinger from a description of the novel inspired by
the actions of the main characters. Characters’ attributes, including such things
as their location, activities and allegiance became the predicates of the planning
domain. The main actions, those that modify characters’ attributes, were repre-
sented as planning operators. The predicates were then used to specify goals and
constraints for individual characters in different story variants and the planner,
DP C , used the operators to generate narratives that satisfied the individual
characters’ goals and constraints (i.e. the perspective of the planner is at the

7
narrative level not the character level). The Goldfinger model that we have de-
veloped is a rich one where the use of constraints enables DP C to generate a
wide range of story variants such as a trajectory consisting of 42 narrative ac-
tions that capture the main outline of the novel (this is the sequence of operators
shown down the centre of figure 4).
Example Narrative Generation
In this example we discuss how our planner DP C generates the narrative shown
in figure 3. This narrative features a goal where Bond triumphs over Goldfinger
by defusing a bomb that Goldfinger was intending to use to irradiate the US
gold reserve, represented with the predicate (defused-bomb bond fort-knox).
If we were to pose this scenario as a planning problem without constraints,
then DP C would simply invoke the base planner with the final goal (in terms of
the algorithm in figure 2 the constraints tree would be empty so control would
pass directly to line 9). A narrative plan would be generated and the output
narrative would be directed towards the goal resulting in a sequence of operators
that contribute directly to achieving the goal. This sequence is shown down the
left hand side of figure 3 culminating in Bond defusing the bomb, represented
by the planning operator (defuse-bomb bond fort-knox).
However, the structure of the narrative can be dramatically changed with
the introduction of constraints which force certain key events to occur. For ex-
ample, the sequence of actions down the centre of figure 3 shows the different
narrative that results when the following constraints are included: at some stage
Bond will beat Goldfinger at golf, represented in the model with the predicate
(won-golf bond goldfinger golf-game); Bond will seduce Goldfingers assistant Jill
Masterson, (seduced bond jill); and before the seduction Bond will beat Goldfin-
ger at cards, (won-cards bond goldfinger card-game) and also receive his mission,
(got-mission bond). These constraints were discussed in section 2 and the corre-
sponding temporal order is shown in figure 1.
Following the algorithm in figure 2, DP C tackles the constraints in order
starting with the earliest. In this example there is a choice of earliest predicates:
(got-mission bond), (won-cards bond goldfinger card-game) and (won-golf bond
goldfinger golf-game), and so DP C formulates these as a disjunctive subproblem
goal using the method of [20]. To do this we introduce an artificial predicate as
the goal of the current subproblem and then for each predicate in the disjunc-
tion we add an artificial operator with: that predicate as the precondition; the
artificial predicate as a single add effect; and no delete effects.
Thus for this example, the goal of the first subproblem is represented using
the predicate (artificial) and three operators are introduced that add it: one
with (got-mission bond) as a precondition; one with (won-cards bond goldfinger
card-game); and one with (won-golf bond goldfinger golf-game). Then the base
planner determines which operator to use to achieve the artificial goal. For this
example, the base planner chooses the artificial operator with the precondition
(got-mission bond) and this is achieved with narrative sequence that takes Bond
to London to receive his mission. At the end of this first iteration of the algorithm
the constrained predicate has been achieved and the other two constrained pred-

8
Fig. 3. Using constraints to shape a narrative (see text for worked example).
icates remain in the temporal order to be tackled on subsequent iterations. The
sequence of operators to achieve (got-mission bond) forms the initial narrative
and DP C then grows the narrative forwards from this point.
Attention now turns to the current earliest constrained predicates in the
temporal order. These are (won-cards bond goldfinger card-game) and (won-golf
bond goldfinger golf-game). Once again DP C poses this as a disjunctive problem
and the state that has been reached at the end of the narrative so far (ie Bond is
in London having received his mission) is used as the starting situation for the
next subproblem. For this subproblem the base planner chooses to next solve
(won-cards bond goldfinger card-game), shown as constrained predicate (2) in
figure 3, with a sequence of operators involving the characters moving to the
venue of the card game and participating in the game. These operators are then
appended to the developing output narrative as it is grown forwards.
Now the earliest predicates in the temporal order are: (seduced bond jill)
and (won-golf bond goldfinger golf-game) and once again these are posed as a
disjunctive problem and the base planner is invoked. As shown in figure 3 the
base planner chooses to solve (seduced bond jill) next with a narrative sequence
that involves Jill changing her allegiance and being seduced by Bond. At this
point there is a single earliest predicate in the temporal order, (won-golf bond

9
goldfinger golf-game), and the base planner outputs a narrative sequence for this
which is appended to the output narrative.
Once all the constrained predicates have been achieved in order, the base
planner is then invoked with the final conjunctive goal, which in this exam-
ple consists of the single predicate (defused-bomb bond fort-knox). Finally the
narrative for this final goal is appended to the output narrative generated so far.
This example demonstrates the way in which the constraints can be used to
support a high level description of a narrative so that it follows the required
dramatic arc. In fact, so long as the set of constraints for a planning instance
accurately reflects the high level control that is required for the narrative then
any plan that is generated that satisfies the constraints (i.e. a valid plan), can
be said to display the required narrative control. This example also shows how
the use of constraints can allow for variation in the narratives that are generated
when we have a partial order over predicates in the domain, since a number of
different variants will satisfy the same set of constraints.
Narrative Segment: introducing conflict
Figure 4 shows a Goldfinger narrative that captures the main outline of the
novel. One way to look at this sequence of events is from the perspective of the
novel’s main theme, the conflict between the protagonist James Bond and his
adversary Auric Goldfinger. As observed by Cheong and Young [8], this type of
conflict, where one characters’ individual goals are the negation of anothers’, is
an important prerequisite for creating suspense in a narrative. Since suspense is
a key feature of Goldfinger, we can use a dimension that places these characters
and the pursuit of their goals in opposition and to view story progression through
the sequence of operators and resulting character situations. Interestingly, the
“trajectories“ along that dimension represent story evolution in a similar way to
“dramatic arcs“ [23, 24].
We can use constraints to introduce conflict into the narrative by imposing
constraints that draw the narrative in directions that set the goal of one of the
central characters against the others. As an example consider the constrained
predicates 8–13 in the lower segment of figure 4. Some of these constraints place
Goldfinger in a position of superiority over Bond (for example, when he has
captured Bond and is attempting to kill him with an industrial laser (10) and
when he has Bond handcuffed to a nuclear bomb (12)), whereas others place
Bond in the superior role and serve to negate Goldfingers goals (for example,
Bond spying on Goldfinger in Switzerland (9) and Bond seducing Pussy Galore
and securing her allegiance (11)).
Narrative Segment: introducing causality
Another way in which the constraints can be used is to promote causality in
the trajectory that is output, for example, to ensure that character behaviour is
justified. As an illustration, consider the constrained predicates numbered 3–5
in figure 4. Here the constraint (won-cards bond goldfinger) is followed by the
constraint (seduced bond jill) and yields a narrative sequence that sets up Jills
betrayal of Goldfinger – she is impressed by Bonds card playing and this causes

10
Fig. 4. Using constraints to shape a narrative which reflects the central theme of the
novel and introduces conflict between Bond and Goldfinger (see text for more details).

11
her to change allegiance – which justifies the subsequent sequence of actions
where Goldfinger has her killed once he becomes aware of her betrayal in order
to assert his position of power.
Narrative Segment: varying pace
We can use the constraints to vary the pace of parts of the narrative. For ex-
ample, consider the constrained predicates numbered 1–3 in figure 4. Bond is
initially in Miami and must travel in order to take part in a card game with
Goldfinger. In the absence of constraints a planner would generate a narrative
which has Bond travelling directly to the venue, with the key action (travel-
to bond miami card-game). However the consequence of introducing constraints
that order (know-of bond goldfinger) to occur in the plan before (enroute bond
miami card-game) which in turn must occur before (won-cards bond goldfinger
card-game) is the addition of the key actions which have Bond and Felix Leiter
meeting in Miami, Felix informing him about Goldfinger and his nefarious activi-
ties, and then introducing detail about Bonds journey from Miami. The inclusion
of the constraints results in the addition of a number of relevant operators which
change the pace by increasing the detail about this characters activities during
this segment of the narrative.
5 Conclusions and Future Work
In the paper we have introduced a novel method for injecting control into au-
tomatically generated narratives for IS. The approach uses features of the plan-
ning representation language PDDL3.0 to encode control knowledge in the form
of constraints and we have developed a novel method for planning with these
constraints. Our results are encouraging and support the hypothesis that these
constraints can be used for narrative control.
Using constraints means that narrative structuring information is specified
independently to the rest of the representation and yields important benefits. For
example, one benefit is that the structuring information can be independently
manipulated and changed without affecting the rest of the representation (this
is much harder if the control information is embedded within pre-conditions
of operators or HTN decompositions). Also for authoring, it appears that the
declarative nature of the constraints makes them easier to express and facilitates
author testing of alternate story variants.
We are currently extending our approach to handle interactivity and in par-
ticular the re-specification of constraints “on-the-fly” in response to user inter-
action. We have identified mechanisms which should support the re-specification
of constraints during an interactive session, suggesting that interactivity could
affect not just the planning domain but also the constraints themselves thus
opening the way for interpreting user intervention at multiple levels.
Acknowledgements This work has been funded (in part) by the European
Commission under grant agreement IRIS (FP7-ICT-231824).

12
References
1. Weld, D.: An introduction to least commitment planning. AI magazine (1994)
2. Riedl, M., Young, M.: An Intent-Driven Planner for Multi-Agent Story Generation.
In: Proceedings of AAMAS. (2004)
3. Aylett, R., Dias, J., Paiva, A.: An affectively-driven planner for synthetic charac-
ters. In: Proceedings of ICAPS. (2006)
4. Bonet, B., Geffner, H.: Planning as heuristic search: New results. In: Proceedings
of the European Conference on Planning (ECP). (1999)
5. Pizzi, D., Cavazza, M., Whittaker, A., Lugrin, J.: Automatic Generation of Game
Level Solutions as Storyboards. In: Proceedings of AIIDE. (2008)
6. Erol, K., Hendler, J., Nau, D.: UMCP: A sound and complete procedure for
hierarchical task-network planning. In: Proceedings of AIPS. (1994)
7. Cavazza, M., Charles, F., Mead, S.: Character-based Interactive Storytelling. IEEE
Intelligent Systems, special issue on AI in Interactive Entertainment (2002)
8. Cheong, Y., Young, R.: A Computational Model of Narrative Generation for Sus-
pense. In: AAAI 2006 Computational Aesthetic Workshop. (2006)
9. Weyhrauch, P.: Guiding Interactive Drama. PhD thesis, School of Computer
Science, Carnegie Mellon University (1997)
10. Magerko, B., Laird, J., Assanie, M., Kerfoot, A., Stokes, D.: AI Characters and
Directors for Interactive Computer Games. In: Proceedings of AAAI. (2004)
11. Mateas, M., Stern, A.: Structuring Content in the Fa¸cade Interactive Drama
Architecture. In: Proceedings of AIIDE. (2005)
12. Riedl, M., Stern, A.: Believable Agents and Intelligent Story Adaptation for Inter-
active Storytelling. In: Proceedings of TIDSE. (2006)
13. Riedl, M.: Incorporating Authorial Intent into Generative Narrative Systems. In:
Proceedings AAAI Spring Symposium on Intelligent Narrative Technologies. (2009)
14. Edelkamp, S., Jabbar, S., Nazih, M.: Large-Scale Optimal PDDL3 Planning with
MIPS-XXL. In: Proceedings of Planning Competition, ICAPS. (2006)
15. Barthes, R.: Image Music Text. Fontana Press (1993)
16. Cavazza, M., Charles, F., Mead, S.: Multi-modal Acting in Mixed Reality Inter-
active Storytelling. IEEE Multimedia (2004)
17. Gerevini, A., Long, D.: Plan Constraints and Preferences in PDDL3. Technical
report, Department of Electronics for Automation, University of Brescia, Italy
(2005) http://www.cs.yale.edu/homes/dvm/papers/pddl-ipc5.pdf.
18. Hoffmann, J., Porteous, J., Sebastia, L.: Ordered Landmarks in Planning. JAIR
22 (2004)
19. Fikes, R., Nilsson, N.: STRIPS: A new approach to the application of theorem
proving to problem solving. Artificial Intelligence (1971)
20. Gazen, B.C., Knoblock, C.: Combining the Expressiveness of UCPOP with the
Efficiency of Graphplan. In: Proceedings of European Conference on Planning
(ECP). (1997)
21. Hoffmann, J., Nebel, B.: The FF Planning System: Fast Plan Generation through
Heuristic Search. JAIR (2001)
22. The Strathclyde Planning Group: VAL: The Automatic Validation Tool for PDDL
including PDDL3 and PDDL+ http://planning.cis.strath.ac.uk/VAL/.
23. Mateas, M.: A Neo-Aristotelian Theory of Interactive Drama. In: AAAI Spring
Symposium on AI and Interactive Entertainment. (2000)
24. Zagalo, N., Barker, A., Branco, V.: Story Reaction Structures to Emotion Detec-
tion. In: Proceedings of ACM Multimedia 2004, Workshop on Story Representa-
tion, Mechanism, Context. (2004)