Pong
- This is work done by Kailash Nadh.
- The game learns to play that video game I played back in the
70s called Pong.
- The input is the Pong game itself, and this connects to the
ball and PaddleInter nets.
- PaddleInter sends activation to the associated PaddleNet.
- Learning is only on between the PaddleNet and the BallNet.
- When the agent plays, the game reads from PaddleNet to move
the paddle.
- So, when a user plays the net associates ball positions with
paddle positions. (It can learn to play badly.)
- It also learns from scratch on its own because when the
paddle misses, the ball passes into the paddle area,
so failure attracts the paddle to the right place.
- It also uses "massively overlapping" CAs, so it can generalise
pretty quickly.
