This project was developed for the AIĀ² Tournament, where the goal was to train reinforcement learning (RL) agents to compete in a 1v1 fighting game similar to Brawhalla. The agents learned to optimize their strategies through self-play and carefully designed reward functions.
š Our agent was awarded the AIĀ² Crown
Our approach combined deep reinforcement learning with strategic hard-coded behavior to ensure robustness in various combat scenarios.
- Proximal Policy Optimization (PPO): Agent trained using a deep neural network with PPO, a leading policy optimization algorithm.
- Multi-Layer Perceptron (MLP) Policy: The agentās decision-making was powered by a deep MLP policy and value network optimized for fighting games.
- Adaptive Movement: The agent learned to approach opponents, execute multi-hit combos, and dodge attacks.
- Survival Mechanism: Hard-coded logic prevented the agent from falling off the platform by detecting edges and adjusting movement accordingly.
- Reward-Based Learning: The agent was rewarded for successful attacks, minimizing damage taken, executing combos, winning matches, and knocking out opponents.
The agent's neural network follows a deep hierarchical structure, with separate architectures for policy learning and value estimation:
This repository incudes the env_final.ipynb notebook which includes the reward functions used, core agent architecture, and the training code. UTMIST_AI2_Env_OFFICIAL.ipynb is the original version of the notebook, with the basic environment configurations and agent class implementations.
This Tournament and most of the code in the notebooks was written and designed by UTMist (University of Toronto Machine Intelligence Team) Members. A huge thanks for creating this exciting challenge!