Learning continuous control policies by stochastic value gradients

TLDR

We present a unified framework for learning continuous control policies via backpropagation. The framework treats stochasticity in the Bellman equation deterministically through exogenous noise, relies on learned models with only real environment observations to reduce compounded errors, and is applied to toy and physics‑based control problems. It yields a spectrum of policy‑gradient algorithms from model‑free to model‑based, and SVG(1) demonstrates simultaneous learning of models, value functions, and policies in continuous domains.

Abstract

We present a unified framework for learning continuous control policies using backpropagation. It supports stochastic control by treating stochasticity in the Bellman equation as a deterministic function of exogenous noise. The product is a spectrum of general policy gradient algorithms that range from model-free methods with value functions to model-based methods without value functions. We use learned models but only require observations from the environment instead of observations from model-predicted trajectories, minimizing the impact of compounded model errors. We apply these algorithms first to a toy stochastic control problem and then to several physics-based control problems in simulation. One of these variants, SVG(1), shows the effectiveness of learning models, value functions, and policies simultaneously in continuous domains.