JSI123/ppo-LunarLander-v2

PPO LunarLander-v2

A Proximal Policy Optimization (PPO) agent implemented from scratch using PyTorch and trained on the LunarLander-v2 environment.

Training

• Algorithm: PPO
• Framework: PyTorch
• Total environment transitions: 999,424
• Parallel environments: 4
• Actor-critic architecture: 64 → 64 hidden units
• Activation: Tanh
• Training phases: 2

Evaluation

Evaluated over 20 episodes for each policy mode.

Metric	Deterministic	Stochastic
Mean reward	35.89	51.22
Standard deviation	127.76	121.16
Minimum reward	-198.93	-132.69
Maximum reward	167.33	191.68

The agent demonstrates learned landing behaviour but remains inconsistent. The evaluation results should be considered alongside the high reward variance.

Files

• ppo_lunarlander_phase2.pt: trained PyTorch model weights
• phase2_evaluation.json: full evaluation results
• training_config.json: training configuration
• replay.mp4: recorded deterministic demonstration episode

Reproducibility

The model uses an actor-critic neural network with separate policy and value networks. The training implementation and notebook will be added separately.

Limitations

The reported results do not establish consistently successful landings. The demonstration video is one selected episode and is not representative of average performance.