General videogame learning with neural-evolution

This paper describes the development of a general learning test, in which an agent's ability to learn to play different games is tested. We used a neuro-evolved agent, which main feature is the use of raw pixels as input, in contrast with common approaches that require some feature extraction defined by an expert. To evaluate the agents we used two games: Pong and Breakout. With these games a cross learning test is used to visualize the knowledge transfer ability of the agents.