Sneks solutions

This repository contains a set of different solutions and studies for the many sneks environments. This has multiple goals:

• Easy to read and explained implementation of various algorithm, from the more classical ones to state-of-the-art methods. The algorithms will be implemented using PyTorch.
• Studies to understand state representation issues in complex visual policies in reinforcement-learning.
• Studies to explore the instability of multi-agent systems, as suggested by this OpenAI post.

Draft of Roadmap

1. Implementation of vanilla DQN and tests on different environments.
2. Basic multi-agent competitive setting: learning by self-play

Ideas

• Perception studies: is the trained policy able to understand a particular scene? Image reconstruction?
• Recurrent policies and partial observability?
• RL fast optimization + evolutionary slow optimization?