|Did you know ...||Search Documentation:|
|Pack logicmoo_nars -- ImportantDocs/riedl-aaai-ss09.txt|
Incorporating Authorial Intent into Generative Narrative Systems
Mark O. Riedl
School of Interactive Computing, Georgia Institute of Techology 85 Fifth Street NW, Atlanta, Georgia 30308, USA tied| @cc.gatech.edu
One of the major themes to emerge in interactive narra- tive research is authorability and authorial intent. With interactive narratives, the human author is not present at run-time. Thus authoring interactive narratives is often a process of anticipating user actions in differ- ent contexts and using computational mechanisms and data structures for responding to the participant. Gen- erative approaches to interactive narrative, in which an automated narrative generation system assumes some of the authoring responsibility, further decouple the haman designer from the participants experience. We describe a general mechanism, called author goals, which can be used by human authors to assert authorial intent over generative narrative systems.
An interactive narrative is an approach to interactive enter- tainment in which a system attempts to tell a story to an interactive participant, such that the user is afforded the op- portunity to make decisions that directly affect the direction and/or outcome of the story. One of the major themes to emerge in interactive narrative research is authorability and authorial intent. Authorial intent is the ability of an au- tonomous interactive system to reflect the intentions of the human designer — also called the human author. Because of participant agency, in an interactive narrative much of a par- ticipants actual run-time experience is influenced by the par- ticipants own actions. Unlike tabletop and live-action role- playing games, the human designer is not present at run-time and cannot make decisions about how the participants expe- rience must be adapted to balance plot coherence and per- ceived participant self-agency. Thus authoring interactive narratives is often a process of anticipating user actions in different contexts and using computational mechanisms and data structures for responding to the user.
An approach to interactive narrative that may mitigate the authoring complexity is generative drama management (or generative experience management (Riedl et al. 2008) for non-dramatic contexts). The generative approach to interac- tive narrative suggests that if authoring branching stories is intractable for human authors then a computer system can
Copyright © 2009, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.
generate story content in response to the actions and deci- sions of an interactive participant. The goal is to have the participants experience become part of an unfolding story. As the participant exerts his or her agency and deviates from the originally intended story, the experience manager in- vokes an automated story generation system to adapt, mod- ify, or re-generate story content. Computers are very use- ful for performing recursive and repetitive tasks. A genera- tive approach to narrative is favorable under circumstances in which there is too much variability for a human designer to foresee all eventualities.
Authoring interactive narrative content is thus a process of instilling a computational system with the ability to make the same decisions that the human designer would make in response to participant actions. That is, the human designers goal is to infuse his artistic vision and authorial intent into a computational system with the tools and data representations at hand. Generative experience management further decou- ples the human designer from the user. That is, not only will the human designer not be present to make decisions at run-time, but, with generative experience management, the human designer is also not responsible for authoring the nar- trative branches that will used to respond to user actions.
In this paper, we describe an attempt to provide mecha- nisms for injecting the designers intent into generative nar- trative systems. The mechanism we describe is called au- thor goals, specialized data structures that are injected into planning-based generative systems that can be used by hu- man authors to indicate preferences in narrative structure.
We consider two types of generative narrative system. The first are narrative generation systems in which the purpose is to automatically produce a non-interactive narrative se- quence. Many narrative generation systems (e.g., (Meehan 1976; Lebowitz 1985; Pérez y Pérez and Sharples 2001; Riedl and Young 2004)) are either based on planning or encapsulate planning-like processes. The second type of generative narrative system are those expressly developed to create real-time interactive experiences. See (Riedl et al. 2008) and (Roberts and Isbell 2008) for reviews of relevant work on interactive narrative systems. Note that two systems in particular, (Young et al. 2004) and (Riedl et al. 2008), ex- plicitly rely on the recursive invocation of a narrative plan- ner. By doing so, these systems build a tree where each child is a re-planned version of the parent narrative plan that han- dles some significant participant action.
Partial-order planning (c.f., (Weld 1994)) is a process of selecting and instantiating actions in a particular temporal order. Plans are comprised of actions. Actions have precon- ditions and effects. Preconditions dictate what must be true in the world for an action to be executed. Effects specify what will be different about the world once the action has been executed. Initially, the root of the search space is an empty plan and the goal state propositions are the only con- ditions that need to be satisfied. When an action (or the goal state) in a plan has a precondition that is not established by a preceding action (or the initial state) a new action is instan- tiated or an existing action is reused to satisfy that precondi- tion or goal. An in-depth discussion of planning is beyond the scope of this paper.
Myers (1996; 2000) explored ways of controlling plan- ners through advice through abstract specifications that are compiled into the planning domain. Advice includes task advice to identify goals and actions to be included, strategic advice to recommend how goals and actions are to be ac- complished, and “evaluational” advice to inform the heuris- tic for overall plan evaluation. Thomas and Young (2006) extend Myers work by creating an environment for human authors to encode preferences through a domain metatheory.
Author goals serve two important purposes. First, author goals constrain the narrative search space such that it is im- possible for a planner to produce a narrative that does not meet certain criteria imposed by the human author. Second, author goals can be used to force complexity in narrative generation.
Technically, author goals are a reformulation of islands for partial-order planners. Islands — a term coined to refer to a technique for controlling the form of solutions generated by planners (Hayes-Roth and Hayes-Roth 1979) — are inter- mediate states in a search space, through which all solutions to a planning problem must pass. In the early days of AI planning research, islands were used to inform the planner as to what valid solution plans should look like, conceptu- ally speeding up the planning process. Potential solutions that do not satisfy each island state description at some point between the initial state and the end state are pruned.
Islands are tools for making state-space search practical for planning purposes. However, many modern planners use plan-space search. See (Weld 1994) for a discussion of the practical advantages of partial-order plan-space search over state-space search. As tools for making planning more prag- matic, islands are not typically necessary for partial-order plan-space search since they can search deeper ply. We use islands as a way for the human user to inject guidance into the narrative generation process and to force the planner to consider more complex action sequences.
Authorial Intent with Author Goals
A narrative generator necessarily operates without a human- in-the-loop. Author goals provide the ability to provide
rough direction for what must occur within solutions. Fur- ther, a generative drama management system will automati- cally produce branches. Because those branches may or may not preserve elements from the root narrative the human au- thor is encouraged to provide additional meta-data. The first meta-data is the outcome. The outcome is a description of the state of the story world after the story is complete. The second type of meta-data is author goals. In terms of autho- rial intent, an author goal indicates that there is a state of the world that must be achieved between the time the narrative starts and the outcome, and that any plan cannot be consid- ered complete unless that world state is at least momentarily true.
Complexifying Narrative Plans
There are many narratives in which states reverse themselves one or more times. For example: a character that begins rich, becomes poor, and finally regains the state of being tich. These phenomena are challenging for planners without some form of guidance. For example, if a planner were given an initial state in which a character was rich and an outcome state in which the character is rich, the planner would simply indicate that there was no problem to solve. Author goals can be used to force the generator to consider substantially more complex plans in which some intermediate state, such as the character becoming poor, must be integrated into the resultant narrative structure.
Incorporating Author Goals into Planning
In our computational representation of narrative, author goals are implemented as a special type of plan step that have preconditions describing the intermediate world state but no effects. Author goals are provided at the time of planner ini- tialization and describe world states that must be achieved at some intermediate time during plan execution. If more than one author goal is given, there can be pre-specified temporal links between them so that author goals must occur in the re- sulting, complete plan in a particular order. In this way, the existence of author goals constrains the space of plans that can be searched by the planner. That is, the planner cannot consider any plan in which the world state described by an author goal will not be achieved during plan execution.
To implement the ability for a planner to act on author goals, we make the following change to the way in which problems are described to the planner. Author goals are specified as sets of state propositions. State propositions, like a planning problem goal, define a set of states in which the given propositions are true. Author goals are specified in the form (set,...set,) where each set, is of the form (author goal,...author goal) and an author goal, is of the form (proposition,...proposition,,). Author goal sets are ordered, meaning that plans must achieve the author goals in the prescribed set order to be considered valid.
The above specifications for author-goals are translated into plan steps and inserted into the initial empty plan. For each author-goal a plan step data structure is created such that the state propositions make up the plan steps precondi- tion list. The plan step has no effects. If there is ordering
character red) (human red) (alive red)
P ( (character wolf) (monster wolf} (alive wolf (character granny) (human granny) (alive granny) (character hunter) (human hunter) (alive hunter} (thing cake) (has red cake) (knows red granny) (knows granny red)) rauthorgoals ((((eaten red)}) (((eaten granny)))))
routcome ((has granny cake)
(:net (eaten red})
(:noet (eaten granny))))}
Figure 1: The Little Red Riding Hood planning problem def- inition.
between author-goals, temporal links are added to the ini- tial plan as well. Partial-order planning algorithms such as those based on (Weld 1994) do not need to be modified fur- ther. These planning algorithms treat the preconditions of the special author-goal plan steps as goals to be satisfied as normal. That is, the planner sees unsatisfied preconditions as flaws and attempts to instantiate an action (or select an existing action) that has an effect that unifies with the un- satisfied condition. The POP algorithm itself does not dis- tinguish between an unsatisfied goal and an unsatisfied pre- condition on an existing action. Generative techniques not based on partial-order planner may require additional modi- fications to the generative algorithm itself to be able to take advantage of author goals.
In this section, we consider two case studies of generative narrative systems in which author goals were essential to their success. Both systems were built on a generative ex- perience management framework described in (Ried1 et al. 2008). The key consideration is that this framework uses partial-order planning technologies to generate narratives.
Little Red Riding Hood
The uses of author goals in a Little Red Riding Hood interac- tive narrative illustrate their necessity in complexifying nar- tative structure. Figure 1 shows the modified problem ini- tialization in a PDDL-like language. An initial state defines characters, character traits, and relevant props and features of the world. The outcome is the goal: Granny has the cake and neither Little Red or Granny are in the state “eaten.” The planning system is also initialized with an action library that describes ways in which the world can change. For exam- ple, characters can give things to other characters and some characters can eat other characters whole.
The author-goals define that a significant feature of gen- eration is that Little Red and Granny should both, at some point, enter the state of being “eaten.” Arguably, the Lit- tle Red Riding Hood domain could not be considered such without Little Red and Granny being eaten and in need of rescue. Note that the initialization parameters do not indi- cate how the author-goals or the outcome are achieved, only that they must be achieved. Author goals are necessary for
1: Red Greet Wolf
(knows wolf red)
2: Red Tell Wolf About Granny (knows wolf red)
(knows wolf granny)
3: Wolf Eat Red
mS . 4: Wolf Eat Granny
(eaten red) a -“ (eaten granny)
Author Goal 1 ws, . Author Goal 2 5: Hunter Kill Wolf
~(alive wolf) ~(alive Nw
6: Red Escape Wolf | | 7: oo Escape Wolf
~(eaten granny) 8: Red Give Granny Cake
~(eaten red) (has granny cake) ~(eaten granny)
Figure 2: Example narrative plan set in the Little Red Riding Hood world.
complexification. Little Red and Granny begin the story in the state of being not eaten and end the story in the state of being not eaten. Without some indication that it is not de- sirable from a story-telling point of view that propositions about these characters states should change dramatically. Without author-goals, a planning algorithm could naively generate the following:
e Red gives Granny the cake e The End.
The author goals given earlier prevent this by forcing the planner to consider substantially more complex plans in which Little Red and Granny become eaten and then are re- stored to not being eaten. See Figure 2 for one possible nar- rative plan that can be generated by respecting the author goals. Boxes are actions or author goals. Solid arrows are causal links — annotations on the plan that capture causal relationships. Dashed arrows are additional temporal con- straints.
Socio-Cultural Awareness Training
The second case study is that of a socio-cultural awareness training prototype developed for the military. This system used an interactive narrative to expose trainees to socio- cultural situations in which dramatic situations unfold to challenge the trainee. In the scenario developed, two mer- chants in a foreign city are involved in a domestic dispute that escalates to violence, eventually involving the trainee acting in the role of peacekeeper. In the training scenario, author goals are used to express authorial intent. The pur- pose of the training scenario is to challenge the trainee by creating dilemma situations where the appropriate course of action to take is not obvious without some deeper socio- cultural situational understanding.
Figure 3 shows the narrative plan. Some causal links are omitted for clarity. In this case study, there are three au- thor goals: (a) the characters are established such that Hasan has acted suspiciously, Saleh has acted unfriendly, and Ali has acted unreliably, (b) a significant incident occurs such as an attack on the marketplace, and (c) the trainee is pre- sented with two (on the surface level) equal possible courses of action, namely that Saleh is falsely accused of causing the incident and that Hasan is accurately accused by an un- teliable character. The first author goal serves the purpose of introducing the characters. This set up stage is causally unnecessary to achieving the outcome, but is considered au- thorially important. The second author goal enforces the constraint that an incident occurs that instigates the final outcome dilemma since it could conceivably arise in other ways. Finally, the outcome state defines the dilemma condi- tions under which the trainee must act. The author goals were necessary because the authorial intentions were ex- tremely hard to encode into the domain itself — the author’s intentions were meta-constraints on the form of the physical action that actually occurs.
The author goal mechanism described in this paper is an at- tempt to enable human authors to inject their preferences, in- tuitions, and requirements into the planning process. Author goals also have the pragmatic side effect that they can force more complexity in narrative plan solutions. In general, au- thor goals constrain the planner to produce narrative plans with particular structures by pruning branches of the plan search space in which plans do not meet the author goals. We believe that enabling the author to inject control into the planning process will become more and more important as narrative systems such as narrative generators or interactive narratives acquire greater autonomy from the human author.
Hayes-Roth, B., and Hayes-Roth, F 1979. A cognitive model of planning. Cognitive Science 3.
Lebowitz, M. 1985. Story-telling as planning and learning. Poetics 14.
Meehan, J. 1976. The Metanovel: Writing Stories by Com- puter. Ph.D. Dissertation, Yale University.
Myers, K. 1996. Advisable planning systems. In Tate, A., ed., Advanced Planning Technology. Menlo Park: AAAI Press.
Myers, K. 2000. Domain metatheories: Enabling user- centric planning. In Proceedings of the AAAI Workshop on Representational Issues for Real-World Planning Systems.
Pérez y Pérez, R., and Sharples, M. 2001. Mexica: A computer model of a cognitive account of creative writing. Journal of Experimental and Theoretical Artificial Intelli- gence 13.
1: Saleh Appear-Unfriendly
(unreliable ali) Author Goal 1
y 4: Hasan Acquire bomb1
2: Hasan Arouse-Suspicion 3: Ali Cry-Wolf \ (unfriendly saleh)
(has hasan bomb1)
6: Hasan Plant bombl
(planted bomb1) AND (armed bomb1)
7: Dud bombl (knows ali (has hasan bomb2))
(incident market) (incident market)
Author Goal 2
8: Ali True-Accuse Hasan
9: Hasan False-Accuse Saleh
(accused saleh) (accused hasan)
Figure 3: Example narrative for socio-cultural training.
Riedl, M., and Young, R. 2004. An Intent-Driven Planner for Multi-Agent Story Generation. In Proc. of the 3rd Int. Conf. on Autonomous Agents and Multi-Agent Systems.
Riedl, M.; Stern, A.; Dini, D. M.; and Alderman, J. M. 2008. Dynamic Experience Management in Virtual Worlds for Entertainment, Education, and Training. /nternational Transactions on System Science and Applications 4(2).
Roberts, D., and Isbell, C. 2008. A survey and qualitative analysis of recent advances in drama management. Inter- national Transactions on Systems Science and Applications 4(2).
Thomas, J., and Young, R. 2006. Author in the loop: Using mixed-initiative planning to improve interactive narrative. In Proceedings of the ICAPS 2006 Workshop on AI Plan- ning for Computer Games and Synthetic Characters. Weld, D. 1994. An introduction to least commitment plan- ning. Al Magazine 15.
Young, R.; Riedl, M.; Branly, M.; Jhala, A.; Martin, R.; and Saretto, C. 2004. An Architecture for Integrating Plan- Based Behavior Generation with Interactive Game Envi- ronments. Journal of Game Development 1.