BahasaMelayuMalaysiaEnglish (UK)


Authors: Lee Mei Ing and Jason Teo



This paper presents the application of deep learning for neuroevolution in General Video Games AI (GVGAI). Single layer neural networks have been always popular in solving problems including games. The focus is then shifted to deep learning which has been proven to solve learning complex and non-smooth functions such as pattern recognition and machine learning. Hence, the extension of deep learning in games has proven that it is also able to solve single games such as board games. However, no one has yet studied deep learning for general game AI. In GVGAI, there are thirty single-player games and each one of them consisted of five levels. In this project, a deep learner is created and it is used to control the agent in the set games. Optimization is applied to optimize the weights in the deep learner. This paper also analyses the performance of the evolved deep learner for game playing by using the set of benchmark games. The deep learner is trained by using the current game state as input and the output as the action to be applied in the game. Experiments are conducted by testing on different optimization algorithms: evolutionary algorithm and differential evolution. The experiments are conducted by testing five controllers including random controller, EA controller, rule-based controller, DE controller and one-step-look-ahead controller. In the result shows that DE controller with five hidden layers are performing equally well as one-steplook- ahead controller which uses heuristic for evaluation. The findings proved that DE deep learner is performing at 100% efficiency compared to the current best
available controller.