Towards Accelerated and Robust Rreinforcement Learning with Transfer Learning
Ruiyong Cao · 2022 International Conference on Big Data, Information and Computer Network (BDICN) · 2022
Transfer learning has been successfully used in some areas in deep learning, such as accelerate training progress in convolutional neural network (CNN). In another area, Reinforcement Learning (RL), time consumption during training is a serious problem due to the dataset. This is why many researchers try to find how to boost up training progress of Deep Reinforcement Learning. Therefore, this paper mainly focus on exploring whether transfer learning will have the same effect on RL as it performed in past CNN models. Thus, this research mainly tests one method of transfer learning, pretrained method, on OpenAI Gym games. In detail, Deep Q Learning was used in playing two Atari games SpaceInvaders and AirRaid, and the experiment for transfer learning was designed to see if what is learned from one game can be adapted to another, this experiment is mainly about utilizing pretrained network in different games and continue training it, including normal training and freezing training. Finally, the conclusion obtained from this research is that transfer learning not only robust training progress to reduce time consumption, but also lead to better performance for another task, potentially presenting the value of transfer learning for training in Deep Reinforcement Learning.