Swarm Tetris: Applying particle swarm optimization to tetris

摘要

This paper investigates the applicability of swarm-based algorithms to the game of Tetris. This work proposes an approach to the problem in which neural network weight values are optimized using a particle swarm optimization (PSO) algorithm. Such an approach has not previously been demonstrated as feasible for Tetris. The reported experimental results show the learning progress of the algorithm, as well as a comparison against a hand-optimized Tetris playing algorithm. The results indicate that the Tetris agents show a continuous improvement over the course of training. Since the experimental focus was on the feasibility of the approach rather than optimizing performance, optimized PSO-based agents were found to be outperformed by the hand-optimized algorithm. However, the playing strategies of the two agents were compared and shown to be similar. The results indicate that a swarm-based approach is feasible, and warrants further investigation.