Multi-task Transfer with Practice

Adapting feedback-driven deep reinforcement learning (DRL) algorithms to real-world problems requires developing robust systems that balance generalization and specialization. DRL algorithms powered by deep neural network function approximation tend to over-fit and perform poorly in new situations. Multi-task learning is a popular approach to reduce over-fitting by increasing input diversity, which in turn improves generalization capabilities. However, optimizing for multiple tasks often leads to distraction and performance oscillation. In this work, transfer learning paradigm Practice is introduced as an auxiliary task to stabilize distributed multi-task learning and enhance generalization. Experimental results demonstrate that the DRL algorithm supplemented with the state dynamics information produced by Practice improves performance.