Playing Mega Man II with Neuroevolution

摘要

The problem of developing Game-Playing Agents provides a controlled environment with varying levels of difficulty in order to test different Artificial Intelligence algorithms. A recently proposed framework for testing such algorithms is called EvoMan and was created based on a classic and challenging game called MegaMan II. In this framework, the agent must defeat a number of different enemies equipped with a diverse set of weapons with different behaviors. This paper follows up the Evoman: Game-playing Competition hosted at the World Conference on Computational Intelligence in 2020 with the objective of finding a general strategy capable of defeating all of the bosses training only on a subset of those. Our approach is composed of manually crafted inputs based on the available sensors fed into a Neuroevolution algorithm composed of a Genetic Algorithm evolving the weights of a Multilayer Perceptron. Our results obtained the first place on the competition and was capable of defeating the entire set of enemies.