Sunday, December 22, 2002

The Turing Tournament

Two very interesting papers on learning in experimental games are posted on the Turing Tournament webste at Caltech. One is an experimental paper by Arifovic, McKelvey and Pevnitskaya, which focuses on the ability of standard learning models to account for the behavior of human subjects in selected finitely repeated games. There are some striking patterns in the human data that standard learning models consistently fail to replicate, such as alternation between the two pure strategy equilibria in the repeated battle-of-the-sexes, and significant cooperation in the repeated Prisoners' Dilemma. The companion paper is by McKelvey and Palfrey, and calls for the development of "strategic learning" models, which allow for the learning not just of stage-game actions but also of repeated game strategies. The Turing Tournament itself is a fascinating attempt to elicit the development of better learning models and I hope that the tragic and untimely death of Richard McKelvey doesn't derail the project.