Michael Buro, Quinn Kybartas, and Santiago Ontañón
The Association for the Advancement of Artificial Intelligence’s 2021 International Conference on Artificial Intelligence and Interactive Digital Entertainment was held October 11-15, 2021. There were three workshops in the program: Experimental AI in Games, Programming Languages in Entertainment, and Strategy Games. This report contains summaries of some, but not all symposia.
Experimental AI in Games
The 2021 Experimental AI in Games Workshop helped to encourage experimentation and discovery in game AI research and game development. This year saw fourteen presentations exploring established subjects such as level and narrative generation, to theoretical and practitioner work. This year also saw the introduction of a “game jam” to attempt to encourage greater practitioner presence within the workshop.
The 2021 Experimental AI in Games (EXAG) workshop is an annual workshop which aims to foster experimental, novel, and unique approaches to game research AI and game development. The workshop was open to submissions from practitioners, researchers and hobbyists, with the goal of creating a dynamic exchange between research and practice. This year saw thirteen accepted papers and one invited presentation, which were presented over the course of two days.
One new feature introduced in the 2021 workshop was the introduction of a “game jam” competition, where non-researchers were able to submit experimental AI games, with the best results being invited as a speaker at the workshop. The turnout of the game jam was minor, having only one resulting presentation, but it nonetheless remains an interesting venue in which submissions to the conference can be made without a corresponding paper, with the peer review phase being held as part of the evaluation of the game jam submissions. Alongside the game jam, EXAG also specifically hosted a practitioner track, welcoming submissions from in-development or released games which make heavy use of unique AI systems. The highlights of the jam and practitioner papers included an open world simulation game for teaching reinforcement learning, and a mobile phone game powered heavily by procedural content generation of worlds, history, and stories.
The majority of EXAG 2021’s submissions were in the research category, and are roughly divided into five categories: Level Generation, Game AI Frameworks, Procedural Content Generation (PCG), Player Modeling and Game Design. Highlights from the level generation works include NLP techniques for generating virtual scenes, the metrics and generation of horror levels, generating settlements in the video game “Minecraft”, and using graph-grammar approaches to dungeon generation. From the PCG track, some highlights include the application of behaviour trees to design generative methods, and the combination of world and story design for roleplaying games. Player modelling papers examined how player models can be used to guide arguments or enable agent manipulation in video games, as well as theoretical work exploring the mechanics afforded by open player modelling. Framework papers focused on co-creative AI driven creation of art and the design of a system for the simple generation of games. Lastly, though theory was present in many of the previous works, there was one purely theoretical paper which explored the application of AI as a gameplay mechanic itself, and also fostered a conversation about AI ethics, and the future of AI in games.
In examining the results of the conference, it can be said this year made more effort to attract practitioner papers, including the introduction of the game jam. In terms of the content, while PCG and level generation are frequent topics in past EXAG workshops, there was a rise in theoretical papers, and a greater exploration of the ethics of AIs use in the game industry. The workshop overall presented both a snapshot of the state of game AI in society, as well as taking steps towards imagining what game AI might look like in the future. The EXAG conference this year was organized by Quinn Kybartas, Mads Johansen and M Charity. This report was written by Quinn Kybartas.
Programming Languages in Entertainment
Programming languages are everywhere in interactive entertainment, often hiding in plain sight – from high-level tools for creating games and generative art, to scripting languages for non-player characters, to level and artifact representation formats, to compilers from high-level designer specifications to runnable code, to game engines driven by language-like rules.
The PLIE workshop aimed to bring researchers and practitioners together who work at the intersection of Programming Languages (PL) and Interactive Entertainment, both broadly construed; especially PL methods applied in game- and art-related contexts. The goal was to bring together researchers across communities, i.e. who participate primarily in one of these communities but have an interest in both areas, to identify convergent lines of work and spark new collaborations.
No formal workshop report was submitted by the organizers.
Strategy Games
The ninth workshop on Artificial Intelligence for Strategy Games was held on October 11, 2021 as a whole-day event at the AIIDE-2021 conference. This report briefly describes the workshop format and its contributions, and provides links to all workshop resources.
The goal of this workshop, which is the ninth in its series, was to bring together AI researchers and programmers from academia and industry, to present and exchange ideas, and to discuss how academia and industry can work together to improve the state-of-the-art in AI for strategy games in general, and real-time strategy (RTS) games in particular. In such games players have to manage economies, build structures and armies, and try to win by destroying all opponents’ buildings. RTS games are interesting from an AI point of view because their decision complexity generated by vast maps, large unit numbers, combinatorial concurrent durative action sets, and limited state observability, precludes solutions based on brute-force search and forces us to consider problem decompositions and abstractions.
This year’s workshop attracted 30 attendees, which is a record. During the one-day event, two peer-reviewed papers were presented, six short invited talks were given on a wide variety of topics, two research projects were described in our “show-and-tell” session, results from this year’s AIIDE StarCraft AI competition 1, and IEEE-CIG μRTS competition 2 were presented, and future research directions were discussed. In particular:
- This year’s peer-reviewed papers focused on RTS game AI. Specifically, the first paper addressed the problem of integrating imitation learning with reinforcement learning to train RTS agents effectively, presented promising results in μRTS, and discussed on-going work on extensions to handle partial observability. The second paper focused on how to inject strategic map information into jump-point search to account for threats, dynamic aspects of the map, and enemy unit movement prediction for fog-of-war for pathfinding. Promising results were demonstrated in StarCraft.
- In the following invited presentation sessions, six authors presented their current work on programmatic strategies which can be represented by short programs humans can read, ongoing work on integrating μRTS into OpenAI’s gym RL framework, combining search and scripts to create high-performance μRTS bots, optimizing traffic signals via Monte Carlo tree search and traffic simulation, training neural network architectures in RL settings for learning tile traversal costs with the help of combinatorial blackbox solvers, and using influence maps with heuristic search to plan sneak-attacks in StarCraft.
- During our Show-and-Tell session, two workshop attendees showcased current work on abstract forward model learning in strategy games, and described the Boulder Dash AI challenge, showcasing the game’s complexity of long-term planning in highly-dynamic environments.
- Finally, during the discussion session, the following challenges were identified as some of the most promising lines of research for the near future:
- How to make it easier to use strategy game AI resources in AI courses? Proposed solutions included using μRTS (an open source abstract RTS simulator developed by Santiago Ontañón that includes a lot of AI system implementations), or using STARcraft, a system developed by Dave Churchill to get started with StarCraft AI development as fast as possible.
- What is the next AI benchmark game? Perhaps a cooperative imperfect information game such as Contract Bridge or a social deduction game such as Among Us, or a complex single-player video game such as Zelda or Boulder Dash?
- How to scale up ML and search for more complex tasks? Attendees suggested to abstract both state and action spaces – and perhaps learn both – and to use hierarchical planning over abstract search/action spaces, learning general concepts first, and with that knowledge to train systems for specific tasks (e.g., first learn kinematics principles, then learn to drive autonomously)
Workshop co-organizer Michael Buro is a professor in the Computing Science Department at the University of Alberta in Edmonton, Canada. Workshop co-organizer Santiago Ontañón a researcher at Google and associate professor in the Department of Computer Science at Drexel University in Philadelphia, U.S.A. This report was written by Michael and Santiago.
More details about the workshop are available at https://skatgame.net/mburo/aiide21ws/.
Biographies
Michael Buro is a professor in the computing Science Department at the University of Alberta
Quinn Kybartas is a researcher and digital maker at McGill University
Santiago Ontañón a researcher at Google and associate professor in the Department of Computer Science at Drexel University