부산대학교 도서관

Using visual observation of environments to identify an ideal action is the problem Reinforcement Learning attempts to solve. Even though several algorithms have used convolutional neural networks, they are not very efficient at learning the representations quickly and generally require large periods of time to converge. Transfer learning has been used as a means to minimize training time and resources in machine learning as it removes the need for a large dataset. This paper describes an approach to implementing transfer learning in actor-critic methods by integrating a pre-trained ResNet50 in the approach to Asynchronous Advantage Actor Critic (A3C). The proposed method is known as ResNet Transfer Learning in Reinforcement Learning (ResTLRL) and it demonstrates that transfer learning can be applied to working with different environments with an improvement of over 68% in terms of maximum rewards when compared to the original implementation on OpenAI Atari benchmarks.