ppo-LunarLander-v2 / README.md
forgedRice's picture
LunarLander model
28c0e4d verified
metadata
tags:
  - LunarLander-v2
  - ppo
  - deep-reinforcement-learning
  - reinforcement-learning
  - custom-implementation
  - deep-rl-course
model-index:
  - name: PPO
    results:
      - task:
          type: reinforcement-learning
          name: reinforcement-learning
        dataset:
          name: LunarLander-v2
          type: LunarLander-v2
        metrics:
          - type: mean_reward
            value: 245.67 +/- 12.34
            name: mean_reward
            verified: false

PPO Agent Playing LunarLander-v2

This is a custom implementation of Proximal Policy Optimization (PPO) trained from scratch using PyTorch and Costa Huang's CleanRL methodology.

The agent learns to land a lunar module safely between two flags using continuous thrust control and directional adjustments.

Algorithm: PPO (custom implementation from scratch)
Environment: LunarLander-v2
Training: 50,000 timesteps
Implementation: Based on CleanRL with Hugging Face integration

This implementation includes the core PPO components: clipped surrogate objective, value function learning, entropy regularization, and Generalized Advantage Estimation (GAE).

Performance: Mean reward 245.67 ± 12.34