Material Pruning Decision Trees in Chess: Reinforcing Pavlovian Behavior by Modeling Human Cognition with Alpha-Beta Pruning

Abstract

Due to the limitations of the brain's computational resources, it is frequently infeasible for humans to consider all potential consequences of an action in arriving at a decision. In a world where AI engines can arrive at optimal outcomes across a myriad of complex scenarios that weren't possible before, studying and modeling the human cognitive process has risen in popularity. In other words, how do humans arrive at the decisions they make given their limited resources? Intuitively, the is answer is that humans must prune their decision trees, but optimal pruning mechanisms that humans may or may not utilize have yet to be formalized. We explore the question if people are efficient in how they use their cognitive resources when they make difficult decisions by studying human cognition in the gamified setting of chess. More specifically, we designed an algorithmic approach and unveiled that chess players adopted depth-limited search as well as a general pruning strategy that involved curtailing the further evaluation of sequences that involved losses of material context in their mental decision trees, scaling with the magnitude of material lost and persisting even when decidedly counterproductive. The Lichess database and Stockfish chess engine are used to automate the process of data collection, aggregation, and evaluation, enabling the generalization of player actions to affirm cognitive theory. Chess players with higher rating and players in looser time constraints were also found to prune by material contexts more accurately in relation to the true values of pieces and implement a more involved depth-limited search than their counterparts, representing a strictly more effective pruning strategy. We ultimately conclude that humans do prune their decision trees by material loss in chess, expanding on the idea of Pavlovian behavioural pruning.