rubenbranco/sokobanalphago

AlphaGo Zero Reinforcement Learning Sokoban Solver

Sokoban AlphaGo Zero (WIP)

Sokoban Solver using DeepMinds AlphaGo Zero system, described by Silver et al 2017, download deepmind unformatted upload available here.

Implementation based on suragnair implementation. Pytorch classification used can be cloned from here.

All board examples come from abelmartin's website.

This is a work in progress and more details about the implementation will be given and the README will be updated as the project moves forward. The objective is to understand how the process works and whether it will work for singleplayer games(aka puzzles).

sokobanalphago's Issues

Generated Boards

Hi Ruben,

you might want to think about trying to train your algorithm on an environment, which generates a new, solvable room for every game. This ensures that your model is not learning the limited number of solutions for the predefined rooms, as pointed out in Imagination-Augmented Agents for Deep Reinforcement Learning
. I developed an environment for OpenAi gym, which does exactly that.
https://github.com/mpSchrader/gym-sokoban

Best,
Max

Quick Question

Have you solved the puzzle yet?

How portable is the AlphaGos algorithm in puzzle games like Sokoban ?