carrom-ai

Carrom_rl is an open-source simulation environment designed to test intelligent agents in the strategic game of Carrom, using reinforcement learning (RL) methods. Built with Python and leveraging Pygame and Pymunk, the simulator provides an interface for developing and training agents to play Carrom, either as single-player or in doubles. In this game, the agent must control the striker to pocket coins, including the queen, while adhering to specific rules such as covering the queen before the last coin is pocketed. The game features a continuous state and action space, with noise added to the actions to simulate real-world uncertainty. As a result, Carrom_rl offers a challenging environment for developing RL agents that can plan and execute strategies under these complex conditions.

In the Carrom_rl environment, the state consists of the positions of the coins on the board and the current score of the player, while the action is represented as a three-dimensional vector containing the striker’s position, the angle of the shot, and the force applied. The game simulates both single-player and two-player modes, with agents required to make decisions based on the evolving state of the board. Actions are noisy, with random disturbances applied to simulate real-world unpredictability. Carrom_rl includes various rulesets for both single-player and doubles formats, allowing agents to compete either against a fixed opponent or cooperate with another agent. This environment is a valuable tool for testing RL agents in a highly dynamic and multi-agent game, providing insights into strategy formation, decision-making, and performance in complex game environments.