Super Mario Bros RL

• [x] Advantage Actor critic [1]

• [x] Parallel Advantage Actor critic [2]

• [x] Noisy Networks for Exploration [3]

• [x] Proximal Policy Optimization Algorithms [4]

• [x] Curiosity-driven Exploration by Self-supervised Prediction [5] (WIP)

• 'Random Network Distillation' pytorch model

• 'Curiosity-driven Exploration' pytorch model

1. Setup

Requirements

• python3.6
• gym-super-mario-bros
• OpenCV Python
• PyTorch
• tensorboardX

2. How to Train

Modify the parameters in mario_a2c.py as you like.

python3 mario_a2c.py

or

python3 mario_ppo.py

3. How to Eval

Modify the is_load_model, is_render parameters in mario_a2c.py as you like.

python3 mario_a2c.py

or

python3 mario_ppo.py

4. Loss/Reward Graph

It use just A2C(PAAC)

It use just ICM and no ext reward.(Curiosity-driven)

References

[1] Actor-Critic Algorithms
[2] Efficient Parallel Methods for Deep Reinforcement Learning
[3] Noisy Networks for Exploration
[4] Proximal Policy Optimization Algorithms
[5] Curiosity-driven Exploration by Self-supervised Prediction