GoExplore-Atari-PyTorch

Implementation of First return, then explore (Go-Explore) by Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, Jeff Clune. The result is a neural network policy that reaches a score of 2500 on the Atari environment MontezumaRevenge.

Content

1. Exploration Phase with demonstration generation
2. Robustification Phase (PPO + SIL + Backward algorithm)

Results

1. Exploration Phase demos with 2500 score
   
   
   
2. Robustification Phase with 2500 score
   
   
   
3. Robustification Phase backward algorithm progress
   
   
   
4. Robustification Phase path