solve_gym

• various methods and reinforcement learning libraries (stable baseline, my own) to solve various openAI gym environments
• note that stable baseline uses a different virtualenvironment. i used autoswitch-venv for this.
• currently i implemented monte carlo (some borrowed here), various TD (SARSA family, Q-learning family), and bellman equations (linear programming, policy & value iterations). each implementation comes with a test example.
• TODO
  • add rounds parameter for policy & value iteration
  • add prioritize replay to DQN
  • add automated tests
  • add more complex algorithms like AC family, PG family, etc. along with example solutions.
  • perhaps change the API a bit so that it looks more like stable baseline's.
  • there's a good NEAT library in Python and perhaps i'd reproduce it for a lot of the examples. (parameter tuning can be quite time consuming though)
  • another general purpose algorithm like NEAT is MuZero, which is very effective at atari games and board games. so perhaps implement that too. (need a beefier machine)