Feature Learning And Transfer Performance Prediction For Video Reinforcement Learning Tasks Via A Siamese Convolutional Neural Network

In this paper, we handle the negative transfer problem by a deep learning method to predict the transfer performance (positive/negative transfer) between two reinforcement learning tasks. We consider same domain transfer for video reinforcement learning tasks such as video games which can be described as images and perceived by an agent with visual ability. Our method directly trains a neural network from raw task descriptions without other prior knowledge such as models of tasks, target task samples and human experience. The architecture of our neural network consists of two parts: a siamese convolutional neural network to learn the features of each pair of tasks and a softmax layer to predict the binary transfer performance. We conduct extensive experiments in the maze domain and the Ms. PacMan domain to evaluate the performance of our method. The results show the effectiveness and superiority of our method compared with the baseline methods.