Adaptive and Multiple Time-scale Eligibility Traces for Online Deep Reinforcement Learning
share
Deep reinforcement learning (DRL) is one of the promising approaches to make robots accomplish complicated tasks. In the robotic problems with time-varying environment, online DRL is required since the methods that directly reuse the stored experience data cannot follow the change of the environment. Eligibility traces method is well known as an online learning technique to improve sample efficiency in the traditional reinforcement learning with linear regressors, not DRL. The one reason why the eligibility traces are not integrated with DRL is because dependencies between parameters of deep neural networks would destroy the eligibility traces. To mitigate this problem, this study proposes a new eligibility traces method that makes it possible to be applied even into DRL. The eligibility traces in DRL accumulate gradients computed based on the past parameters, which are different from that computed based on the latest parameters. Hence, the proposed method considers the divergence between the past and latest parameters to adaptively decay the eligibility traces. Instead of that divergence directly, Bregman divergences between outputs computed by the past and latest parameters, which are computationally feasible, are exploited. In addition, inspired by the replacing eligibility traces, a generalized method with multiple time-scale traces are newly designed. In benchmark tasks on a dynamic robotic simulator, the proposed method outperformed the conventional methods in terms of the learning speed and the quality of the tasks by the learned policy. A real-robot demonstration verified the importance of online DRL and the adaptability of the proposed method to the time-varying environment.
READ FULL TEXT