Why to do Research on Games?

After a long time, here I go again... I gave up on promising to write posts on a regular basis because I will always have some other things to be concerned with. I hope to not give up posting... To "re-start slow", I will not discuss something specific to Machine Learning with equations and so on. Today I will try to answer a simple question: Why to do Research in Games?

------

My Master's thesis was about player modelling in the game Civilization IV (I will write a post about it someday). Basically one wants to learn human players' preferences in the game from their behaviour. The benefits of player modelling in the game are obvious, ranging from predicting when a player will quit to customizing the game for maximizing players entertainment. However, in the early days of my research, I was not sure whether this research was relevant in a greater sense. Is research in games "just" to make games better? When I started to look for an answer I quickly realized it is beyond that. Since it was not obvious for me that time, and I know some researchers who think this is "silly", I will try to answer why it is relevant to do research in games.

First, as I state in my homepage,  mutiagent systems have several applications such as; auctions, negotiation, security and surveillance. Such approaches deal with uncertainty in the true state of the world, uncertainty about future events and uncertainty in the decisions made by other agents. It is particularly difficult to evaluate proposed ideas, as full-scale deployment and testing is costly in such environments. Games, however, offer a self-contained domain with an easier path for testing and evaluation, where all the rules of the environment are know, success and failure are well defined, etc.

Fairclough et al. [2001] for example argue that “computer games offer an accessible platform upon which serious cognitive research can be engaged” [1], while Laird and van Lent [2000] suggest that computer games are the perfect platform to pursue research into human level AI [2].

There are also several applications of techniques developed in the games domain and that were later applied to different problems. Probably the most famous success of AI in games was DeepBlue, the first computer to beat the Chess World-Champion (Garry Kasparov). Later, "in 1999, IBM developed the Deep Computing Institute built on its experience with Deep Blue. The institute aimed to explore and solve large, complex technological problems through deep computing—using large-scale advanced computational methods applied to large data sets. Deep computing provides business decision makers with the ability to analyze and develop solutions to these very difficult problems."  This has been applied to Data Mining, Financial Modeling, Molecular Dynamics, etc [3]. Another example in the realm of board games are techniques developed to solve Checkers [4] that were later used by a bioinformatics company for biological computations.

We can state the same for techniques developed to generate AI for Poker. Two examples are (1) Conterfactual Regret Minimization [5], a technique to obtain a Nash Equilibrium that was later used in the design of diabetes treatment policies [6] and; (2) DIVAT [7] and Imaginary Observations [8], evaluation metrics used to create Poker Bots but that were recently used to study online gambling [9]. These techniques were used to model psychologic profiles of gamblers, showing that those who play online have different perceptions about their skill (are optimistic).

In summary, the take-home message is that doing research in games is relevant, cool and funny! One uses games to evaluate ideas that may be used in other domains. It is not just about "making games better". There are certainly many other cases where techniques developed to games were later applied to "real-world problems". I hope this small selection was enough to convince you about the relevance of such research. In fact, if you know any other case as those listed here, please let me know. Post in the comments and I will be very thankful!



References

[1] Fairclough, C., Fagan, M., Mac Namee, B., and Cunningham, P. (2001). Research Directions for AI in Computer Games. In Proceedings of the 12th Irish Conference on Artificial Intelligence and Cognitive Science, AICS, pages 333–344.

[2] Laird, J. E. and Lent, M. V. (2001). Human-Level AI’s Killer Application. AI Magazine, 22(2):15–26.
 
[3] IBM. Deep Blue - Transforming the World. Available at: http://www-03.ibm.com/ibm/history/ibm100/us/en/icons/deepblue/transform/ Last access: April, 27 2014.

[4] Solving Checkers, showing that perfect play by both sides leads to a draw  Jonathan Schaeffer, Neil Burch, Yngvi Björnsson, Akihiro Kishimoto, Martin Müller, Robert Lake, Paul Lu, and Steven Sutphen. (2007). Checkers Is Solved. Science, 317(1518).

[5] Martin Zinkevich, Michael Johanson, Michael Bowling, and Carmelo Piccione. (2008). Regret Minimization in Games with Incomplete Information. In NIPS.

[6] Katherine Chen, and Michael Bowling. (2012). Tractable Objectives for Robust Policy Optimization, In NIPS.

[7] Martin Zinkevich, Michael H. Bowling, Nolan Bard, Morgan Kan, Darse Billings. (2006). Optimal Unbiased Estimators for Evaluating Agent Performance. In AAAI, pages 573-579.

[8] Michael H. Bowling, Michael Johanson, Neil Burch, Duane Szafron. (2008). Strategy evaluation in extensive games with importance sampling. In ICML, pages 72-79.

[9] T. L. MacKay, Nolan Bard, Michael Bowling, D. C. Hodgins. (2014). Do pokers players know how good they are? Accuracy of poker skill estimation in online and offline players. Computers in Human Behavior, 31, pages 419-424.