ML-Agents-SnowballFight-1vs1

What is ML-Agents-SnowballFight-1vs1?

ML-Agents-SnowballFight-1vs1 is a sophisticated multi-agent reinforcement learning environment built using Unity ML-Agents Toolkit. It simulates a competitive 1v1 snowball fight where agents must learn both offensive and defensive strategies to succeed. The environment demonstrates practical applications of multi-agent reinforcement learning in competitive scenarios.

Implementation Details

The model utilizes PPO (Proximal Policy Optimization) with sophisticated observation and action spaces. The neural network architecture features 2 hidden layers with 512 units each, trained with a batch size of 2048 and buffer size of 20480. The implementation includes self-play mechanisms for continuous improvement, with team changes every 200,000 steps and a window of 10 for opponent selection.

• Ray-cast based perception system with 33 rays for environment sensing
• Discrete action space with branched actions for movement and shooting
• Self-play training with ELO rating system starting at 1200
• Reward shaping based on hit success and time penalties

Core Capabilities

• Complex spatial reasoning through ray-cast observations
• Strategic decision making for shooting and movement
• Health and cooldown management
• Advanced opponent tracking and avoidance