Policy Optimization as Online Learning with Mediator Feedback

Policy Optimization (PO) is a widely used approach to address continuous control tasks. In this paper, we introduce the notion of mediator feedback that frames PO as an online learning problem over policy space. The additional available information, compared standard bandit feedback, allows reusing samples generated by one estimate performance other policies. Based on observation, propose algorithm, RANDomized-exploration via Multiple Importance Sampling with Truncation (RANDOMIST), for regret minimization in PO, employs randomized exploration strategy, differently from existing optimistic approaches. When space finite, show under certain circumstances, it possible achieve constant regret, while always enjoying logarithmic regret. We also derive problem-dependent lower bounds. Then, extend RANDOMIST compact spaces. Finally, provide numerical simulations finite and spaces, comparison baselines.