MVPA and GA Comparison for State Space Optimization at Classic Tetris Game Agent Problem

Abstract

Tetris is one of those games that looks simple and easy to play. Although it seems simple, this game requires strategy and continuous practice to get the best score. This is also what makes Tetris often used as research material, especially research in artificial intelligence. These various studies have been carried out. Starting from applying state-space to reinforcement learning, one of the biggest obstacles of these studies is time. It takes a long to train artificial intelligence to play like a Tetris game expert. Seeing this, in this study,  apply the Genetic Algorithms (GA) and the most valuable player (MVPA) algorithm to optimize state-space training so that artificial intelligence (agents) can play like an expert. The optimization means in this research is to find the best weight in the state space with the minimum possible training time to play Tetris with the highest possible value. The experiment results show that GAs and MVPA are very effective in optimizing the state space in the Tetris game. The MVPA algorithm is also faster in finding solutions. The resulting state space weight can also get a higher value than the GA (MVPA value is 249 million, while the GA value is 68 million).