Additions to MCB

[cgc]

• Adding UI to permit creating a new agent, training an agent more, hand-tuning an agent, or setting hyper-parameters of an agent.
• Making things a little more deterministic, particularly in a way that's tweakable via the UI.
• Consider other ways of training agent besides policy gradients + RL. overview
  • Cross-entropy method. papers on CEM: 0, 1, keras-rl CEM implementation
  • Policy gradient with continuous action space. post 0
  • TRPO? Q-learning? Actor-Critic?

[cgc]

Should consider using deeplearn.js instead of synaptic, which has GPU support. https://deeplearnjs.org