Playing Tetris with deep reinforcement learning

Abstract

Tetris is a hard game to learn due to its random environment, large state space, and need for a long-term strategy. The offline version of the game is shown to be NP-hard. The state-of-the-art approach, CBMPI, is a hybrid algorithm based on evolution algorithms and policy iteration. It uses manually crafted features and achieves 51 million lines cleared in an average game.

In recent years, deep reinforcement learning (DRL) has achieved outstanding performance with Atari and Go games. An initial attempt by Stevens and Pradhan (2016) to use deep reinforcement learning to play Tetris was unsuccessful. The objective of this thesis is to explore the potential of DRL with Tetris games. We started with a baseline algorithm that uses a quadratic reward function and standard Q-learning framework. We experimented with linear-reward to discourage risky moves, harder games to reduce training time, and expected updates to reduce volatility in training. The combination of the three achieves 52-fold increase in performance over the baseline algorithm. With minimal fine-tuning and limited training time (6 hours), our final model achieves an average of 60357.7 pieces survived and an average score of 40163. Our experiments show that deep reinforcement learning has tremendous potential to create a high-performing Tetris player.