frallebini / deep-ga

An implementation of a genetic algorithm aimed at evolving the parameters of a neural network to play Atari games

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Deep GA

Frostbite Space Invaders

We train a convolutional neural network to play two Atari 2600 games from the Arcade Learning Environment, Space Invaders and Frostbite. The model, as in DQN, receives raw pixels in input and outputs action values. However, differently from DQN and other popular approaches, it is trained with a gradient-free, population-based method belonging to the family of genetic algorithms, which allows us to achieve interesting results after training the network on a single desktop and with CPU only.

See report.pdf for details and src for the implementation.

About

An implementation of a genetic algorithm aimed at evolving the parameters of a neural network to play Atari games


Languages

Language:Python 100.0%