A unified framework to control estimation error in reinforcement learning

Yujia Zhang, Lin Li, Wei Wei, Yunpeng Lv, Jiye Liang

Abstract: Highlights•A framework that can combine any Actor–Critic estimation methods is proposed.•Two algorithms and their variants are introduced in the framework.•Elegant weighting methods are designed for more accurate value estimation.•Estimation errors and bias upper bounds are analyzed for different methods.•State-of-the-art performance achieved by proposed methods on multiple benchmarks.