Crazy Tetris: A Reinforcement Learning Approach to Adversarial Tetris

Our project is to implement an agent that learns to play Tetris in an adversarial environment. The Tetris game was invented by Alexey Pajitnov in 1985. The version we use is a 2-dimensional game that consists of a board with a fixed size and a sequence of bocks of different sizes. Whenever the game provides a new block, the player has a certain amount of time steps to place the block on the board, on top of the blocks that have already been placed before. As soon as the block is placed, a new block is provided, and so on. If the new block creates a line across the width of the board, that line disappears and typically the score goes up. The game ends when the board is filled, which means that no new block can be placed on the board without the height of the highest column being larger than the height of the board. There are seven types of blocks (also called tetrominoes), that are shown in Figure 1. The distribution of blocks is unknown to the player and can be fixed or it can change according to the configuration of the board. In the latter case, the game is adversarial if the distribution of blocks assigns lower probability to blocks that would get the player a higher score if he or she was to place them in a optimal way. In our setting we considered the possibility that we might face such an adversarial environment.