Enhancement of CNN-based 2048 Player with Monte-Carlo Tree Search

Abstract

In this study, we developed computer players for a single-player stochastic game 2048 using an existing neural-network evaluation function and a version of Monte-Carlo tree search. We applied the Monte-Carlo softmax search (MCSS) algorithm, with some modifications in order to adapt it to the stochastic game, and designed six methods of controlling the search algorithm. We evaluated the MCSS players in an exhaustive manner and also conducted longer experiments for two MCSS players by changing the number of simulations per move. Our MCSS player achieved an average score of 533 542 under the limit of 2000 simulations per move. This result was better than Expectimax players that used the same evaluation function.