How to beat Atari Pong in 1 Day on CPU without CNN


#1

I know the Google Deep Mind team has done it before with CNNs but I wanted to see if I could get similar results without using Convolutional Neural Networks.


My architecture / approach
Model-Free Q-Learning
Normalized raw pixels as input layer
2 Hidden RELU layers
Controller actions as output layer
90,000 frame memory of State,Action,Reward tuples
Batch training every 30 game
Stochastic game play every 10 game

Results
Starts to achieves consistent wins around 900 games
Achieves 90% win rate around 1200 games