Multi-source Transfer Learning for Deep Reinforcement Learning

Deep reinforcement learning has obtained impressive performance in challenging tasks in recent years. Nevertheless, it has important limitations such as long training times and the number instances that are needed to achieve acceptable performance. Transfer learning offers an alternative to alleviate these limitations. In this paper, we propose a novel method for transferring knowledge from more than one source tasks. First, we select the best source tasks using a regressor that predicts the performance of a pre-trained model in the target task. Then, we apply a selection of relevant convolutional kernels for the target task in order to find a target model with similar number of parameters compared to the source ones. According to the results, our approach outperforms the accumulated reward obtained when learning from scratch in 20.62% using lower parameters (about 56% of the total, depending on the specific game).