Conference Round-Up: IEEE CIG 2015

The format of the Conference Round-Up posts is largely driven by Tommy Thompson’s visits to AI/games conferences across the world as part of his research activities. However, Tommy can’t attend every conference for a multititude of reasons.  But that shouldn’t mean we have to miss out on important updates in the field and the big talking points from conferences.  With that in mind, we welcome Chong-U Lim: a PhD candidate at MIT, who is delivering the conference update on IEEE CIG 2015.  You can learn more about Chong-U at the bottom of this piece.

This September saw the annual iteration of the IEEE Conference on Computational Intelligence and Games (CIG) 2015, held from August 31 – Sep 2, 2015. The conference was located at a fantastic venue and location – the Tayih Landis Hotel in the city of Tainan, Taiwan and jointly hosted by strong contingent of six different academic institutions. CIG is the premier annual event for researchers applying computational (CI) and artificial intelligence (AI) techniques to games, which are regarded widely as one of the best platforms for studying AI. This year saw the conference’s continuing trend in pushing the latest advancements in CI/AI in topics such as the ever-popular Monte-Carlo Tree Search (MCTS), player modeling, content generation, board games, evolutionary computation, and more. In particular, an increasing amount of attention in the realm of general video game playing (GVGP) – with a keynote by Prof. Simon Lucas from the University of Essex, a tutorial conducted by Dr. Diego Perez, also from the University of Essex, the 2nd iteration of the CIG General Video Game Playing competition. In this round up, we shall cover all this and other interesting takeaways from what is continuing to be one of our favorite conferences for AI and Games.

Monte-Carlo Tree Search

A term that’s becoming almost synonymous with the conference itself, Monte-Carlo Tree Search (MCTS) was yet again a popular research topic. As pointed out in last year’s CIG round-up, MCTS is not extremely new in the CI/AI research community, but what is interesting is the ever-expanding application domains that it is being applied and showing to be effective in. Spreading across two tracks, we saw MCTS for exploration of game tree / state space problems, modeling player preferences and learning behaviors, creating new types of opponents, and also for developing agents for general video game playing. As many different variants of MCTS exist, including ways to adjust heuristics, parameters, and augmenting it with other techniques, the talk on “Investigating MCTS Modifications in General Video Game Playing” provided a great overview on how each of them influence its performance.

Content Generation

The content generation track is always informative and exciting. While procedural content generation aims to lift take some weight of the human’s figurative shoulders we saw how combining human computation and content generation was an effective method for generating more aesthetically pleasing and enjoyable levels (“Human Computation for Procedural Content Generation in Platform Games”). One of the strong motivations is generating content that better suits or adapts to players to provide them with more compelling experiences, whether its wall-building behaviors mimicking professionals in StarCraft (“Creating Efficient Walls Using Potential Fields in RTS Games), dynamic difficulty (“A Data-Driven Approach for Online Adaptation of Game Difficulty) or for emergent storytelling (Making Sense of Emergent Narratives: An Architecture Supporting Player-Triggered Narrative Processes). It is great to see how CI/AI techniques are continuing to aid both the creative design process and game playing experiences of users.

Spelunky

Okay, there wasn’t a single track dedicated to Spelunky (… for now.) But in line with our discussion on player-adaptive content generation and the introduction of the SpelunkBots AI Toolkit at last year’s CIG presented by Daniel Scales and our very own Tommy Thompson, it is worth pointing out some recent research developments in this area. As a game with decidedly more complex levels and mechanics, it is a challenge in both designing and completing levels. The paper “Player-Adaptive Spelunky Level Generation” tackles this problem by adapting to players by dynamically adjusting levels. The authors also show interesting results and takeaways such as the implications of making levels less difficult for players, which may not always be a good thing. Nonetheless, it is really continuing research interest into Spelunky and SpelunkBots as a platform for AI/CI research.

General Video Game Playing

Three years ago at the 2012 Dagstuhl Seminar (see round-up of the Dagstuhl 2015 here), the first proposals of developing a general video game language for the purpose of general video game playing were put forth. A year later in 2013, we saw the introduction of the Python Video Game Description Language (VGDL), a python runtime and programming language enabling the creation of simple 2-dimensional (2D) arcade and puzzle games. And last year saw the inaugural General Video Game Playing AI (GVGAI) 2014 competition where competitors created AI controllers using a Java port of the VGDL language (javaVGDL) that could play a multitude of different games that varied in game mechanics, win/lose conditions, and other challenging aspects of video games. At CIG2015 this year, an entire track dedicated to this rapidly evolving research area was present. Papers presented included “Neuroevolution for General Video Game Playing” and “EnHiC: An Enforced Hill Climbing Based System for General Game Playing.” But beyond game playing, there is also the subject of general video game generation, and the paper “Towards Generating Arcade Game Rules with VGDL” is one that seeks to address the challenge of whether VGDL can be used to generate complete arcade games using an evolutionary approach. It was also the recipient of the Best Paper Award at CIG2015 (congratulations!). As we saw earlier, content generation poses challenging research problems and generating game rules is no exception. Overall, all these demonstrate how GVGP and generation are fast becoming a critical research area for CI/AI in games.

Player Modeling

Seizing this opportunity to talk a little bit about my own research, the player modeling track this year saw the application of CI/AI to developing better systems and tools to better understanding players and their behaviors in games. The paper “Predicting Player Disengagement and First Purchase with Event-Frequency Based Data Representation,” outlines an approach to predict a player dropping off or making their very purchase within the game, two key aspects of commercial games. The paper “Toward Avatar Models to Enhance Performance and Engagement in Educational Games,” presented interesting results of the impact of virtual representations on player performance, engagement, and enjoyment in games. It showed how virtual representations such as avatars, which while non-physical and deemed non-“real,” can be empirically shown to have an effect on players. Finally, my very own talk was on “Understanding Players’ Identities and Behavioral Archetypes from Avatar Customization Data,” which demonstrated how a player’s physical-world identity characteristics, such as their reported gender, can be computationally modeled and predicted through their behaviors in virtual environments. These behaviors could also be systematically categorized in groups that differed based on differences between virtual and physical-world identities.

Competitions

One of the most interesting aspects of CIG is the competitions track. This year saw the several familiar competitions such as the StaftCraft AI competition, Fighting Game AI competition, the Geometry Friends Game AI competition, a as mentioned, the General Video Game AI Competition (GVGAI). We also saw the Human vs. Computer-GO Competition and an invited competition of U-Generation Islands of Adventure (E-game) Competition. The results and summaries for each of these competitions are available at the CIG2015 Competitions page, but we wanted to highlight a few interesting snippets that occurred.

StarCraft

The StarCraft Competition, organized by Kyung-Joong Kim, Sehar Shahzad Farooq, In-Seok Oh, and Man-Je Kim from the Cognition and Intellignece Lab in Sejong University had particularly notable results. In the past, two main observations had been made: (1) humans outplay AI consistently and (2) the Protoss traditionally outperformed Terrans, who outperformed the Zerg.  Entries in previous years have always had entries that were either Protoss or Terran. This year, however, not only were there Zerg entries, but the top-three performing AI-bots and winners were Zerg! This is made even more notable as the top three AI-bots were newcomers to the competition.

GVG-AI

The GVG-AI competition, organized by Diego Perez, Spyridon Samothrakis, Julian Togelius, Tom Schaul, and Simon M. Lucas, made its second appearance at CIG2015.  The number of entries (77) had increased almost more than seven-times from last year (10). The performance of the controllers also saw marked improvements, with the Sample MCTS controller falling from 2nd last year down to 19th. In fact, for the first time across all GVGAI competitions (that included the GECCO’15 competition), the winner was a Genetic Algorithm (GA)-based controller. These collectively reiterate the openness and challenge of this research problem.  A second aspect of the GVG-AI competition that is probably worth mentioning is the steadily increasing database of games created with VGDL. Several of these games feature interesting game mechanics, rules, and win conditions that really highlight the broad spectrum of possibilities (and overall challenge) that a controller needs to consider. For instance, take one of the validation set games, “Wait for Breakfast” as an example. While most games often require players to move from start to goal, or battle enemies, or avoid cars, etc., the only way to win in “Wait for Breakfast” was to do nothing. A player had to wait at a table for breakfast, and if would lose if any movement occurred before the food-server had arrived with the food. It was a entertaining, yet thoughtful, example of how robust the VGDL and GVG-AI framework is (which now has a total of 60 unique games to play.) Stay tuned to more competitions, including plans for a learning track (i.e., no forward model) and procedural generation track.

Closing Remarks

We hope that this round-up has provided you as readers with some insight into the CIG conference and the landscape of CI/AI-related research in games that are ongoing by the extremely talented and enthusiastic bunch of people all around. As games continue to grow as an industry, the growth of CI/AI-related research will continue to affect and influence the design and development of these applications, and we have notable influence on other application domains related to AI, too.

But wait! There’s more…

It is worth pointing out that news for the 2016th IEEE Conference on Computational Intelligence and Games is beginning to trickle out. Keeping with the tradition of rotating continents for the conference venue, CIG2016 will be held the beautiful Santorini in Greece. In fact, the first Call for Papers has recently been announced.

Call for proposals just announced! We highly recommend that you take a look at and keep close tabs at the Official CIG2016 Website!

 

References

Papers

  • Frederik Frydenberg, Kasper Andersen, Sebastian Risi, and Julian Togelius.
    Investigating MCTS Modifications in General Video Game Playing
  • Caio Freitas de Oliveira and Charles Maderia. (2015)
    Creating Efficient Walls Using Potential Fields in RTS Games
  • Haiyan Yin, Linbo Luo, Wentong Cai, Yew-Soon Ong, and Jinghui Zhong. (2015)
    A Data-Driven Approach for Online Adaptation of Game Difficulty.
  • Simon Chauvin, Guillaume Levieux, Jean-Yves Donnart, and Stephane Natkin. (2015)
    Making Sense of Emergent Narratives: An Architecture Supporting Player-Triggered Narrative Processes.
  • David Stmer, Tobias Günter, and Mike Preuss (2015)
    Player-Adaptive Spelunky Level Generation
  • Hanting Xie, Sam Devlin, Daniel Kudenko, and Peter Cowling (2015)
    Predicting Player Disengagement and First Purchase with Event-Frequency Based Data Representation
  • Dominic Kao and D. Fox Harrell (2015)
    Toward Avatar Models to Enhance Performance and Engagement in Educational Games
  • Chong-U Lim and D. Fox Harrell (2015)
    Understanding Players’ Identities and Behavioral Archetypes from Avatar Customization Data

Presentations


 

About Chong-U Lim

ChongUChong-U Lim is a Ph. D. Candidate at the Imagination, Computation, and Expression (ICE) Laboratory in the department of Electrical Engineering and Computer Science at MIT. His current research focuses on developing Artificial Intelligence techniques, systems, and applications for studying virtual representations and player behaviors in videogames. While technically virtual, they are capable of reflecting phenomena that occurs in the real world and thus provide a rich, interactive, and compelling medium for better understanding the values, preferences, and phenomena encompassing individuals and society as a whole. More on his research is available at http://people.csail.mit.edu/culim.

Enjoying AI and Games? Please support us on Patreon!
Tommy Thompson Written by:

Tommy is the writer and producer of AI and Games. He’s a senior lecturer in computer science and researcher in artificial intelligence with applications in video games. He’s also an indie video game developer with Table Flip Games. Because y’know… fella’s gotta keep himself busy.

Comments are closed.