Guest Editorial: Computational Narrative and Games
Post on 27-Jan-2017
Embed Size (px)
IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 6, NO. 2, JUNE 2014 93
Guest Editorial:Computational Narrative and GamesI. INTRODUCTION
Narrative is central to our culture, to the ways we communi-cate, and, many have argued, to our cognition itself. AI researchhas engaged with narrative deeply, for instance, in the work ofSchank et al. on story understanding  and generation , inthe narrative intelligence movement of the 1990s , and, morerecently, with the communities surrounding the workshops onComputational Models of Narrative  and Intelligent Narra-tive Technologies .Video and computer games, although not required to be nar-
rative by their nature, have also connected to narrative conceptsin important ways. Most computer games make at least somepretense of having a story, even if only as backstory barely sug-gested in gameplay. The seminalDonkey Kong , for example,could easily have been presented as a purely abstract formalpuzzle, similar to Tetris . But instead, it is framed as thequest of a hero (Jumpman, who developed into the famousMario) attempting to rescue a damsel in distress from the gamestitular villain. One of the earliest computer games, Adventure(also known as Colossal Cave ), had exploration and actionsin its original version, but achieved its full success when DonWoods made a stronger connection to familiar stories by im-porting the tropes of Tolkeinesque fantasy into the game. Thatgame spawned entire genres, including interactive text adven-tures such as the Infocom games and their graphical children,and point-and-click adventure games such as Myst . It wasinfluential in the development of the entire field of electronicliterature.Narrative is also an increasing focus of contemporary AAA
games, so much so that the Game Developers Conference nowhas a dedicated (and quite popular) narrative summit to focuson storytelling practice within video games. This expanded in-terest is driven both by the game industry seeking to expand itsmarket from core gamers to nontraditional audiences, and byplayers who increasingly judge games in terms of their narra-tive qualities. Games such as Heavy Rain  and The WalkingDead  have demonstrated that there are considerable audi-ences for big-budget games that can provide an interactive ex-perience of cinematic storytelling. Individually produced indiegames (such as Gone Home , Dear Esther , and Story-teller ) have also shown how aspects of narrative can informgame design in radical and powerful ways.This special issue presents diverse problems, approaches,
and viewpoints in the area of AI for computational narrativefor games. While this TRANSACTIONS focus is specifically ongames, the problems and promises of computational narrativeare not limited to gaming. Video games provide a compellingdomain with clear cultural relevance and a variety of interesting
Digital Object Identifier 10.1109/TCIAIG.2014.2325879
challenges, but they are only one domain for computationalnarrative. Others include story generation, story understanding,literary and dramatic art, and interactive documentary. Tech-niques developed in any of these domains may, of course, proverelevant for another.
II. COMPUTATIONAL MODELING OF NARRATIVE
Most of the work presented here involves problems of nar-rative generation, broadly construed. This might involve theab initio creation of narratives, the dynamic instantiation ofhuman-authored nonlinear narratives, or the interactive selec-tion of narrative elements. Narrative generation techniques varyin the ways they deal with the tensions between the goals of nar-rative coherence (and story quality more broadly), the ability forhuman designers to exercise authorial control over the storiesthat are produced, and the problems of adapting an unfoldingnarrative to unexpected player actions.Planning-based methods are popular for narrative generation
because they allow explicit modeling of the goals and causalstructure of the story itself, but also of the goals of the au-thor and the reasoning of the reader/player; modeling these ele-ments may prove critical to the generation of very high-qualitystories. However, interactivity introduces challenges to the on-line application of planning-basedmethods within games.Whenplayers are given the freedom to take a range of actions withina game environment, it is a challenge for a system to commit toa storyline in advance; unanticipated player actions may breakplan-based structures that underlie a generated storyline. Sim-ulationist approaches (also known as emergent narrative) seekto address this by modeling the story world explicitly as a dis-tributed simulation in which characters pursue their goals in-dependently of any central controller. These approaches allowmaximum freedom to the player at the cost of reduced authorialcontrol and narrative coherence.Instead of creating a story from whole cloth, one might at-
tempt to guide the player through a prespecified space of pos-sible stories toward the best possible story, given the choicesmade by the player. This technique is often referred to as dramamanagement. In drama management (an approach relevant togames, but not restricted to them), the space of possible storiesis often represented as a set of plot points or beats, together withconstraints on their application or an explicit directed graph ofallowable transitions between them. The drama manager usesits knowledge of the structure of the space of possible stories tochoose events within the story world that will lead to better sto-ries, given the choices made by the player. The metrics used toevaluate the relative merits of potential storylines may be basedon authorial goals, a model of the players preferences, or somegeneral theory of story quality.
1943-068X 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
94 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 6, NO. 2, JUNE 2014
III. RESEARCH AND TRENDS IN THIS ISSUE
The papers presented here explore a range of approaches tocomputational narrative. They extend and combine the tech-niques above, as well as others. Some describe systems that gen-erate or manage a story within a running game, while others lookat problems of offline generation or of user experience issuesduring the interaction. Looking across the papers, a number ofimportant topics and contrasts emerge.Social interaction, and the computational modeling of social
interaction in interactive drama in particular, has been a topic ofincreasing interest in recent years. Two papers presented hereexplore social simulation as both a game mechanic and a driverof narrative.Both papers seek to develop reusable social simulations that
can be used in multiple games/stories while providing the op-portunities for authors to guide the simulations in desired direc-tions.In Social story worlds with Comme il Faut, McCoy et al.
discuss the issues of making CiF a general, reusable socialphysics engine, as well as its use in the social puzzle gameProm Week . Another system, Versu , was the basisfor several interactive fiction games, particularly comedies ofmanners. It is the topic of VersuA simulationist storytellingsystem by Evans and Short. Starting from the work of JaneAusten, Evans and Shorts focus is on the explicit modelingof social practices and norms. In their approach, these normsare used by characters not only to select actions, but also toevaluate and model one another and respond accordingly. Thesocial situations modeled by both systems (whether in highschool or at a Regency tea party) are familiar but also height-ened; they are contexts in which the social dimensions ofthe situation are clearly important ones. Both systems definehigh-level social practices or norms that can differ as the sit-uation varies and that characters can approach differently de-pending upon their traits.The work presented in this issue includes some papers fo-
cused on the story or content level of narrative (the level ofwhat happens?) and some papers focused on the level of dis-course (how it is told?). Research on the underlying storyworld is represented by A computational model of narrativegeneration for surprise arousal by Bae and Young. They opera-tionalize narratological models of surprise by embodying themin a planning-based story generation algorithm that explicitlymodels narrative based on causal centrality in a storys struc-ture and validate their computational model with user studies.At the level of discourse, however, one must decide how tonarrate the events of the story once they have been selected.This is the problem examined in Automated story selectionfor color commentary in sports by Lee et al.While the eventsthat happen during a sports event are determined by the ongoingplay, commentators have some freedom to narrate. Lee et al.ssystem learns a mapping from the internal state of a baseballvideo game to a catalog of stories and uses this mapping to dy-namically suggest stories for inclusion in the running commen-tary of the game.
Skald: Minstrel reconstructed by Tearse et al. reports onreconstructing an influential 1993 system (Scott Turners Min-strel ) to be able to better understand it and to allow furtherexperimentation with the system. The reconstruction engageswith the history of the field, issues of preservation, and ques-tions of how results can be duplicated; it also significantly re-vises our view of Minstrel. Another focused investigation, thisone looking specifically at dialog, interface, and its influenceon user engagement and enjoyment, is offered in Designinguser-character dialog in interactive narratives: An exploratoryexperiment by Endrass et al. Varying the way that players in-teract using speech and text, the authors found a preference to-ward more continuous interaction (along the lines of Faade) over explicit pauses for input, as in classical text adven-ture games.As we described above, approaches that employ drama
management guide a player through an established spaceof possible stories rather than creating stories on the fly. InPersonalized interactive narratives via sequential recommen-dation of plot points, Yu and Riedl present an algorithm thatprovides advice to a player or a reader about how to proceedthrough a branching narrative. Their user studies indicate thatthe advice produced by their approach leads to improved andmore personalized player experiences. In Lessons on usingcomputationally generated influence for shaping narrativeexperiences, Roberts and Isbell, Jr., describe a drama managerthat uses techniques taken from influence theory in socialpsychology to guide players through an experience; their exper-iments found that players did not sense a loss of control whenthis system was operating, while the resulting story experienceswere more aligned with the authors goals. In A supervisedlearning framework for modeling director agent strategies ineducational interactive narrative, Lee et al. discuss a dramamanager that adapts stories to players in an educational gamingcontext. Their system uses machine-learning techniques toselect actions for an in-world agent who serves to direct theplayer. The results were positive, not only for players narrativeexperiences but also for their learning outcomes.Story generation is not limited to traditional techniques such
as planning and simulation, however, and many novel tech-niques have been explored. One such technique is analogy. InShall I compare thee to another story?An empirical studyof analogy-based story generation, Zhu and Ontan presenta study of the Riu system. Riu uses a computational model ofhuman analogical reasoning not only as a technique for gen-erating stories, but also as a sophisticated storytelling devicewithin a given story. Structuralist techniques such as story gram-mars are another important class of narrative generation tech-niques. In Analysis of ReGEN as a graph-rewriting system forquest generation, Kybartas and Verbrugge use context-sensi-tive graph-rewriting rules to automate the generation of side-quests for large-scale games such as Skyrim . Sidequestsare an interesting case of narrative generation for games in that,while they have a narrative structure, they can be generated non-interactively since the enactment of the plot points of the questis the explicit responsibility of the player. One particularly inter-
IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 6, NO. 2, JUNE 2014 95
esting aspect of this work is the attempt to measure the qualityof the stories generated by the system through the use of met-rics such as the number of distinct paths through the quest or theuniqueness of actions with a given quest.Looking across these papers, a number of trends can be seen.
One is the attempt to find principled methods for evaluatingstory generators. Human subject studies based on questionnairedata are common, but the story metrics of ReGEN, and therational reconstruction method of Skald, which provides ananalog of the practice found in the sciences of independent re-production of results by other labs, are interesting approaches.Another is the importation and operationalization of theories ofhuman performance, Bae and Youngs work on Prevoyant, orZhu and Ontans use of structure mapping theory and forcedynamics.
Considerable progress has been made in AI for computa-tional narrative. The deployment of sophisticated AI techniquesin indie games such as Storyteller  and PromWeek , andmainstream commercial game platforms like Versu , markan important step for the field. But considerable work remainsto be done.A significant amount of current work in AI and narrative
seeks to provide authoring tools and assistance to writers andgame designers. In some cases, the research is meant to replacerather than supplement human narrative production. Such workis often undertaken, however, with no or minimal consultationof the experts who are meant to be assisted or emulated by thesystems being developed. While interesting developments canresult, there is the risk that work will proceed along lines that re-late to AI techniques but are aesthetically unproductive. Thereis also the possibility that the tools that are produced are diffi-cult to deploy beyond the AI community or even the originalproject team. If improving authorship tools is a serious goal, orif the goal is providing narratives that are similar to human-pro-duced ones, researchers would do well to involve human do-main experts. Automated story selection for color commentaryin sports provides one example in which this was done effec-tively in the evaluation stage.Current evaluation practices have been imported from
humancomputer interaction (HCI), psychology, and softwareengineering. Thes...