A3C for Kung Fu in OpenAI Gym

This repository contains code for training an A3C agent to play Kung Fu MasterDeterministic-v0 environment in OpenAI Gym.

Key Features:

• Implementation of the A3C (Asynchronous Advantage Actor-Critic) algorithm for multi-agent training.

• Preprocessing pipeline for Kung Fu observations using the PreprocessAtari wrapper.

• Environment batching for parallel interaction with multiple environments.

• Evaluation of the trained agent on single episodes.

• Video recording and visualization of the agent's gameplay.

• Train the agent for 3000 episodes and periodically show the average agent reward during training.

  Environment: https://gymnasium.farama.org/environments/atari/kung_fu_master/

video.mp4

Requirements:

Python 3 PyTorch NumPy OpenAI Gym tqdm

Additional Notes:

The script currently trains 10 agents in 10 parallel environments. You can modify these numbers in the number_environments and EnvBatch class. The reward scaling (batch_rewards *= 0.01) is optional and might need adjustment depending on your environment and training dynamics.