Multi-source Domain Adaptation Based on Data Selector with Soft Actor-Critic

Multi-source domain adaptation (MDA) aims to transfer the knowledge learned from multiple-sources domains to the target domain. Although the source domains are related to the target domain, the difference of data distribution between source and target domains may lead to negative transfer. Therefore, selecting the high-quality source data is conducive to mitigate the problem. However, the existing methods select the data with uniform criteria, neglecting the variety of multiple source domains. In this paper, we propose a reinforced learning Data Selector with the Soft Actor-Critic (DSAC) algorithm for MDA. Specifically, the Soft Actor-Critic (SAC) algorithm has two Q-value Critic networks, it can better judge the performance of the data. Select the data in multi-source domains to migrate with our target domain, and use the difference in loss both before and after the model to determine the quality of the data and whether it is retained. Extensive experiments on the representative benchmark demonstrate that our method performs favorably against the state-of-the-art approaches.