Tetris Reinforcement Learning Agent

A Deep Q-Network (DQN) agent trained to play Tetris using Stable-Baselines3 and tetris-gymnasium.

Overview

The agent uses a custom reward function (V3) with GA-optimized feature weights to encourage line clearing. It observes board state as a feature vector (column heights, holes, bumpiness) and learns via DQN over 1.5 million timesteps.

Reward shaping includes:

• Line-clear bonuses (scaled by number of lines cleared, with a Tetris bonus for 4 at once)
• Penalties for aggregate height, holes, and bumpiness (weights from genetic algorithm optimization)
• Small per-step survival reward
• Game-over penalty