Proceedings of the Annual Meeting of the Cognitive Science Society

What can a machine learning simulation tell us about human performance in a complex, real-time task such as Tetris¬ó? Although Tetris is often used as a research tool (Mayer, 2014), the strategies and methods used by Tetris players have seldom been the explicit focus of study. In Study 1, we use cross-entropy reinforcement learning (CERL) (Szita & Lorincz, 2006; Thiery & Scherrer, 2009) to explore (a) the utility of high-level strategies (goals or objective functions) for maximizing performance and (b) a variety of features and feature-weights (methods) for optimizing a low-level, onezoid optimization strategy. Two of these optimization strategies quickly rise to performance plateaus, whereas two others continued towards higher but more jagged (i.e., variable) plateaus. In Study 2, we compare the zoid (i.e., Tetris piece) placement decisions made by our best CERL models with those made by the full spectrum of novice-to-expert human Tetris players. Across 370,131 episodes collected from 67 human players, the ability of two CERL strategies to classify human zoid placements varied with player expertise from 43% for our lowest scoring novice to around 65% for our three highest scoring experts.