Keywords: Reinforcement Learning, Black-Box Optimization, Hypothesis Testing, Coordinate Ascent, Block Coordinate Ascent, Random Search, MDP
Abstract: This work develops a novel black box optimization technique for learning robust policies for stochastic environments. Through combining coordinate ascent with hypothesis testing, Hypothesis Driven Coordinate Ascent (HDCA) optimizes without computing or estimating gradients. The simplicity of this approach allows it to excel in a distributed setting; its implementation provides an interesting alternative to many state-of-the-art methods for common reinforcement learning environments. HDCA was evaluated on various problems from the MuJoCo physics simulator and OpenAI Gym framework, achieving equivalent or superior results to standard RL benchmarks.
One-sentence Summary: This work develops a novel black box optimization technique that combines coordinate ascent with hypothesis testing to learn robust policies for stochastic environments.
Supplementary Material: zip