Neuroevolution for General Video Game Playing

Abstract

General Video Game Playing (GVGP) allows for the fair evaluation of algorithms and agents as it minimizes the ability of an agent to exploit apriori knowledge in the form of game specific heuristics. In this paper we compare four possible combinations of evolutionary learning using Separable Natural Evolution Strategies as our evolutionary algorithm of choice; linear function approximation with Softmax search and e-greedy policies and neural networks with the same policies. The algorithms explored in this research play each of the games during a sequence of 1000 matches, where the score obtained is used as a measurement of performance. We show that learning is achieved in 8 out of the 10 games employed in this research, without introducing any domain specific knowledge, leading the algorithms to maximize the average score as the number of games played increases.