DISTORTION RISK MEASURE-BASED DEEP REINFORCEMENT LEARNING

Mainstream reinforcement learning (RL) typically focuses on maximizing expected cumulative rewards. In this paper, we explore a risk-sensitive RL setting where the objective is to optimize the distortion risk measure (DRM), a criterion better reflecting human risk perception. We parameterize the action selection policy by neural networks and propose a novel policy gradient algorithm, DRM-based Policy Optimization (DPO), along with its accelerated variant, DRM-based Proximal Policy Optimization (DPPO), to address deep RL tasks with DRM objectives. DPO integrates three coupled recursions operating at different timescales to estimate gradient components and update parameters simultaneously. Our experiments provide numerical results across diverse scenarios, demonstrating that our proposed algorithms outperform the existing baselines under the DRM criterion.