Cross-Game Generalization Approaches for General Video Game Playing using Deep Reinforcement Learning

This paper presents a number of novel approaches to help facilitate Deep Reinforcement Learning (DRL) for Neural Networks based agents in the domain of General Video Game Playing (GVGP). Using common processing methods, the NN can retain a fixed predetermined input and output shape while having the flexibility to play a variety of games. Training can also be made more efficient via reward normalization. Our results show that these modifications negligibly impact the learning time and performance of the model when testedupon games in the GVGAI framework. Furthermore analysis of these approaches within the parameters of the GVGAI learning competition was also performed. This research work attempts to use the adaptations presented to train an agent thatwill compete in the CEC2019 GVGAI Learning Competition.