Reproduction of "Curiosity-driven Exploration for Deep Reinforcement Learning" with PyTorch

This is a PyTorch reproduction of the ICML 2017 paper "Curiosity-driven Exploration for Deep Reinforcement Learning." It uses A2C rather than A3C. It only allows evaluation on the SuperMarioBros environment right now. Look at the original Tensorflow implementation here.

To Train

python3 a2c_devel.py --recurrent-policy

Requirements:

• Python 3.6
• Numpy
• Gym
• Pytorch >=0.4.0
• Matplotlib
• OpenCV
• Baslines
• gym-super-mario-bros

Acknowledgements:

• Credit to @pathak22 For their publicly available original implementation
• Credit to @baselines for templates for the environment wrappers
• Credit to @sadeqa for the template for ProcessFrameMario() wrapper
• Credit to @ikostrikov for A2C, GAE, PPO and inspiration for plotting code implementation