AI Conference Round-Up: IEEE CIG 2014

August saw the IEEE conference on Computational Intelligence in Games (CIG) take place in Dortmund, Germany for 2014.  For those outwith academic circles, CIG is one of the biggest academic conference on AI applied within video games.  The emphasis of methods adopted is fairly broad, with work on algorithm design and experimentation, tools analysis and construction and data mining of player habits and behaviour.  The research presented at CIG is often interested in how games can be used for algorithm design and experimentation.  This often due to the quality of domains that video games can introduce for AI research.

Personally this is one of my favourite conferences to attend: it carries a wide variety of research that often explores a wide range of areas.   This can range from the more ‘traditional’ explorations to quirky personal projects that carry some scientific merit.  So for the remainder of this piece, we highlight the key topics and presentations that caught my interest, as well as what the community in general seems focussed upon.

A Panorama of CI/AI in Games

A tutorial held on the first day of the conference was dedicated to a recent publication by Georgios Yannakakis and Julian Togelius.  This publication (cited and linked to below) is a panorama of all AI research being conducted in games.

A view of how AI interacts in games from the perspective of the player and their interactions. (Yannakakis and Togelius, 2014)
A view of how AI interacts in games from the perspective of the player and their interactions. (Yannakakis and Togelius, 2014)

Why is this so important?  Well, we are rapidly approaching 10 years since the first CIG symposium was held in the UK, not to mention that the AAAI conference AIIDE holds its 10 year anniversary this year.  As a result, there is a significant body of research that has been conducted and the notion of ‘AI research in games’ has exploded: problem domains, methodologies, applications have all arisen, disappeared or thrived within this last decade.  In fact, it is often hard to keep up with how fast this field has changed. Speaking as someone who left the AI and games research community for a few years to explore other pursuits, it takes a lot of reading and networking to get a feel for where everyone is and what are the big talking points.

If you are interested in getting a feel for where we, as a community, are right now – and if you’re reading this, you should – then I would advise reading the paper.  It will prove valuable not only for aspiring young academics, but established researchers as well.

Monte Carlo Tree Search

If there was a year to hail as the point where Monte Carlo Tree Search (MCTS) had grabbed the imagination of the community, then 2014 would be it.  MCTS is nothing new in the games research world and has been around now for a few years: my first experience with it was applied to Klondike Solitaire in 2009, prior to its successes in Computer Go in 2011.  Not only did we see whole sessions of the conference dedicated to MCTS applications such as StarCraft (Justesen et al., 2014) and a continuation of research in Go, but we also saw work that appeared outside of the main MCTS track.  Work by Spyridon Samothrakis appeared in the session that I chaired which was focussed on a horizon planning approach which fit games with continuous physics such as Lunar Lander (Samothrakis et al., 2014).

The GVG-AI Competition

The results are in!  After discussing the topic of general video game playing on this site (in two parts), we finally saw this work come to a head as the first bout of the GVG-AI competition was completed. Perhaps unsurprisingly, MCTS once again proved to be a big deal. as it dominated the competition and the higher rankings.  However, this is not the complete story and there were some surprises when considering the effectiveness of MCTS and where this competition will go in the coming years.     Keep an eye out for a future post as I return to discuss the competition at length: how well my agent performed and the rationale for my design choices behind it.

With Fate Guiding My Every Move…

Spelunkbots

Finally, I can talk a little bit about something I was doing while at CIG.  In fact, the main reason I was there was to discuss a paper of mine that was developed by myself and ex-student Daniel Scales (Scales and Thompson, 2014).  This is the first time anyone has attempted to do this with Spelunky, but we now have a working build of the game that allows developers to write their own bots in C++.  We are only just getting started with this work and I will be talking more about this in the near future.

References

Yannakakis, G.N., and Togelius, J., (2014): A Panorama of Artificial and Computational Intelligence in Games . IEEE Transactions on Computational Intelligence and AI in Games.

Justesen, N., Tillman, B., Togelius, J., and Risi, S., (2014)
Script- and Cluster-based UCT for StarCraft. CIG 2014

Samothrakis, S., Roberts, S., Perez, D. and Lucas, S.M., (2014)
Rolling Horizon methods for Games with Continuous States and Actions. CIG 2014.

Scales, D. and Thompson, T. (2014) SpelunkBots API: An AI Toolset for Spelunky. CIG 2014.

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.