evolvable agent

[honzo0481]

tonight after seeing how well some of the randomly initialized agents played before being trained i think some kind of genetic algorithm / tournament style training could be interesting. A minimal version might work like this:

• 2 agents with randomly initialized NNs (lets say 2 layers: 1 dense, 1 sigmoid, but lots of architectures could work) play a game.
• The winner is cloned to the losers position.
• A random sample of the clones weights are randomly updated.
• Rinse and repeat.