Evolving HyperNetworks for Game-Playing Agents

This work investigates the evolution of indirectly-encoded neural networks through a hypernetwork approach. We find that for some Atari games, a hypernetwork with over 14 times fewer parameters, can compete or even outperform directly-encoded policy networks. While hypernetworks perform worse than directly encoded networks in the game Frostbite, in the game Gravitar, the approach reaches a higher score than any other evolutionary method and outperforms complicated deep reinforcement learning setups such as Rainbow.