A Case for Validation Buffer in Pessimistic Actor-Critic
This addresses error accumulation in reinforcement learning for robotics, but it appears incremental as it builds on existing pessimistic actor-critic methods.
The paper tackles error accumulation in critic networks using pessimistic temporal difference objectives by proposing the Validation Pessimism Learning (VPL) algorithm, which uses a validation buffer to adjust pessimism levels and reports improvements in sample efficiency and performance on locomotion and manipulation tasks.
In this paper, we investigate the issue of error accumulation in critic networks updated via pessimistic temporal difference objectives. We show that the critic approximation error can be approximated via a recursive fixed-point model similar to that of the Bellman value. We use such recursive definition to retrieve the conditions under which the pessimistic critic is unbiased. Building on these insights, we propose Validation Pessimism Learning (VPL) algorithm. VPL uses a small validation buffer to adjust the levels of pessimism throughout the agent training, with the pessimism set such that the approximation error of the critic targets is minimized. We investigate the proposed approach on a variety of locomotion and manipulation tasks and report improvements in sample efficiency and performance.