Inventing the Future of AI for Games: Lessons from EMPath

The space of Narrativity has had the benefit of being
explored for centuries. Technology, however, introduces
a new factor to pioneer by expanding the possibilities
of both the experiencing and the telling of a story.
Advancements of AI provide a multitude of techniques
for the process of authoring and the actual authoring of
stories in games. This process, however, is not so straight
forward, and has proved to be extremely challenging in
practice. Narratives in games, although sharing similar
narrative qualities of its medium predecessors, deliver
a highly interactive experience, making games a matter
of new media and new analysis. The purpose of this talk
is to demonstrate and give an idea of what the contrast
is between the theories and models created in academia
and the practice of industry in using technology to express
compelling and believable experiences.
Introduction
With emphasis on the advancements of believability in
areas of sound and scene, the advancements in the believability
or complexity toward intelligent interactions
within commercial video games are comparably lacking.
Certain games aim to leave the story up to the user’s
imagination, while others reduce user agency to create
an artistically inspired story, and still, others are focused
on procedurally creating a story that has both high user
agency and dramatic significance. Arguments that support
story games demonstrating high levels of intelligence
claim that with the incorporation of Drama Managing,
Multi-Agent Systems, Reactive Planning, and
other sub-areas of AI will come a new frontier in experiencing
and telling stories reducing traditional conflicts
(Mateas 2001). On the other hand, game developers in
industry are resistant against these ideas due to the complications
that come along with adding such methods.
Efforts in commercial games have been more successful
in making the most out of scripted stories, maximizing
from sound and scene, and sacrificing certain qualities
in order to strengthen others. In addition to surveying
the areas that are currently researched in academia and
discussing games developed through research, there will
be a demonstration of the EMPath project. EMPath is a
prototype sized, adventure game that utilizes the Declarative Optimization-Based Drama Manager (DODM).
This prototype game, developed at UC Santa Cruz, is a
real-time playable game that uses the DODM architecture
and has been tested on over 100 users. The purpose
of this demo is to exhibit a novel AI-based approach to
interactive storytelling, as well as provide a concrete example
of the challenges in the design process. Figure 1
is a screen shot of the dungeon map in the original EMPath
game (Sullivan 2009).
Figure 1. The 5x5 map world of EMPath.
Drama Management
A drama manager (DM) monitors an interactive experience,
such as a computer game, and intervenes to shape
the global experience so that it satisfies the author’s expressive
goals without decreasing a player’s interactive
agency. Most research work on drama management has
proposed AI architectures and provided abstract evaluations
of their effectiveness. A smaller body of work has
evaluated the effect of drama management on player experience,
but little attention has been paid to evaluating
the authorial leverage provided by a drama management
architecture: determining, for a given DM architecture,
the additional non-linear story complexity a DM affords
over traditional scripting methods. This poster will propose
three criteria for evaluating the authorial leverage
of a DM: 1) the script-and-trigger complexity of the DM
story policy, 2) the degree of policy change given changes
to story elements, and 3) the average story branching
factor for DM policies vs. script-and-trigger policies
for stories of equivalent quality. Figure 2 illustrates the
decision-tree representation approach in evaluating authorial
leverage. For preliminary studies in evaluating
complexity of the drama manager, thousands of partial
stories were generated and used to train and test decision
trees using the J48 algorithm implemented in Weka, a
machine-learning software package.1 Partial stories (the
independent variable) are represented by a set of boolean
flags indicating whether each plot point and DM action
has happened thus far in this story, and, for each pair of
plot points a and b, whether a preceded b if both happened
(Chen 2009).
Figure 2. Zoomed-in view of a decision tree that has
learned from the DM.
Lessons from EMPath
Gains for creating a more intelligently interactive story
need to be proven through subject testing and other types
of evaluation. In particular, there are two metrics that
need to be established: one to show overall improved
experience, and one to show the lightening of authorial
burden. These approaches are demonstrated in previous
experiments with EMPath; however, the prototype
games that are feasible for research have difficulty demonstrating
significant results due to the scale of these
smaller games. Other challenges that arise from implementing
and evaluating AI systems are as follows (Chen
2009):
• Authoring: Incorporating an AI system creates the
burden of domain understanding, forcing an author
to break traditional habits in authoring.
• Evaluation Measures: Story qualities must be
mathematically represented in order to show improvement
or to encourage better interactions.
• Player Modeling: AI systems often depend on
predicting and anticipating user actions and motivations.
The system would need to model human
tendencies mathematically.
• Simulation: Experiments designed to run offline,
according to the system’s user model are needed
to provide authorial leverage and sanity checks for
story evaluation comparisons.
• User Testing: Users need to notice a difference in
their overall experience when using the AI system
versus not using the AI system. Experiments need to be carefully designed to show improvement in the
user data.
• Game Design: The game must be designed to be
able to demonstrate such differences. In general,
games that are created for research purposes are often
not expansive or large enough to show significant
results.
• Trail Blazing: Finally, a great challenge in creating
AI systems is that evaluation measures and user
study approaches have not been rigorously tested
for these purposes
Dimensions of Narrativity & Interactivity
Taking another look at Narratology, establishing a more
complete understanding of narrative may help ameliorate
some of the challenges in designing and testing AI
systems. If there is a better understanding of the objectives
that an AI system is aiming toward, then the space
of story that it creates may be more easily evaluated. For
instance, reducing the space of possible stories by fixing
the ontological variations in an interactive story space
reduces the variety of experiences, but results in a more
focused output. With a fixed and linear diegetic universe,
both the author’s artistic vision is maintained and, as a
result, the dramatic significance of the vision. Research
can begin by finding ways of delegating types of discourse
actions and performing them on a fixed set of plot
points contingent on the actions of the user in trying to
optimize user agency under those conditions.
Figure 3. Adaptations of Ryan’s 8 narrative dimensions.
By establishing narratives as a relationship among scalar
properties, Marie-Laure Ryan’s dimensions of narrativity
gives a solid means to compare and analyze interactive
narratives (Ryan 2006). For Ryan, her dimensions
are more to establish conditions for the mark of the narrative.
For the purpose of the discussion, a variation of
Ryan’s dimensions will be used analyze the experiential
variations that result from interactivity in narratives.
Figure 3 visually illustrates the analogous dimensions.
The new adaptation, instead of being a model for narratives,
will serve as a model for interactive experiences.
The four dimensions are: temporal (an axis to situate
time), event space (an axis for delineating the occurrences
in a story space), foci (the experiential views from
intelligent existents or perspective), and discourse (an
axis for determining the means of how a story is told).
For further explanation of the relationship between these
dimensions, it helps to “fix” or ignore one or more of the
dimensions.
Conclusion
Overall, this presentation will deliver a broad understanding
of the ways in which AI can integrate with
interactive stories and create a diversity of experiences
and outcomes. In contrast to commercial games, the process
of interactive storytelling provides insights into the
endeavors of universities and research institutions pioneering
the area through advancements in AI. Additionally,
there will be a live demo of the EMPath project in
conjunction to a discussion of the difficulties and challenges
in the counter-intuitive design process of a story
that utilizes concurrent technologies of AI. Ultimately,
this demo will be a case study towards gaining a deeper
understanding of the challenges to be faced before what
is possible in storytelling can be made practical through
the intersections of the arts, sciences, industry, and academia.
Notes
1http://www.cs.waikato.ac.nz/ml/weka/
References
Chen, Nelson, Mateas (2009). Evaluating the Authorial
Leverage of Drama Management. AAAI Spring Symposium,
Interactive Narrative II.
Chen, Sullivan, Nelson, Wardrip-Fruin, Mateas (2009).
Intelligent Interactive-Stories: Theory versus Practice.
Game Developers Conference, San Francisco, CA.
Mateas (2001), A preliminary poetics for interactive
drama and games. Digital Creativity, vol 12, number 3.
Ryan (2006). Narrative, Media, and Modes: Avatars of
Story. University of Minnesota Press.
Sullivan, Chen, Mateas (2009). Abstraction to Reality:
Integrating drama management into a playable game
experience. AAAI Spring Symposium, Interactive Narrative II.