Neural network ensembles for video game AI using evolutionary multi-objective optimization

摘要

Recently, there has been an increasing interest in game artificial intelligence (AI). Game AI is a system that makes the game characters behave like human beings that is able to make smart decisions to achieve the target in a computer or video game. Thus, this study focuses on an automated method of generating artificial neural network (ANN) controller that is able to display good playing behaviors for a commercial video game. In this study, we create neural-based game controller for screen-capture of Ms. Pac-Man using a multi-objective evolutionary algorithm (MOEA) for training or evolving the architectures and connection weights (including biases) in ANN corresponding to conflicting goals of minimizing complexity in ANN and maximizing Ms. Pac-man game score. In particular, we have chosen the commonly-used Pareto Archived Evolution Strategy (PAES) algorithm for this purpose. After the entire training process is completed, the controller is tested for generalization using the optimized networks in single network (single-net) and neural network ensemble (multi-net) environments. The multi-net model is compared to single-net model, and the results reveal that neural network ensemble is able learn to play with good strategies in a complex, dynamic and difficult game environment which is not achievable by the individual neural network.