Learning to Play Tetris from Examples

Abstract

We model a Tetris player using Artificial Neural Network (ANN). In contrast to most previous works which learned weights of predefined heuristic functions, our model learns the actual actions i.e., given a board state and a tetrominoe, where the piece should be placed on the board. The problem is formulated as a classification problem where the model learns to associate a board state with fruitful actions. We compare an ANN player with a random player and a Best First Search (BFS) player. The random player and the BFS player provide baselines for uninformed search and informed search, while the ANN player learns from the examples extracted from BFS runs and its performance lies between both baselines. We observe that generalisation appears to be very hard from the nature of the game itself. Fitting Tetrimino pieces together demands an exact match. This means we cannot expect similar solutions to closely related patterns. In this paper, we explore the problem from a soft computing perspective, present our experimental design and provide a critical discussion of the results of our experiment.