Deep Radial-Basis Value Functions for Continuous Control

TL;DR

This paper introduces deep radial-basis value functions (RBVFs) for continuous control in reinforcement learning, enabling accurate action-value maximization and universal function approximation, leading to improved performance over baselines.

Contribution

The paper proposes deep RBVFs with RBF output layers, extending DQN to continuous actions and demonstrating their effectiveness and universality in value function approximation.

Findings

RBF-DQN outperforms value-function-only baselines.

Deep RBVFs accurately approximate maximum action-values.

The approach is competitive with state-of-the-art actor-critic methods.

Abstract

A core operation in reinforcement learning (RL) is finding an action that is optimal with respect to a learned value function. This operation is often challenging when the learned value function takes continuous actions as input. We introduce deep radial-basis value functions (RBVFs): value functions learned using a deep network with a radial-basis function (RBF) output layer. We show that the maximum action-value with respect to a deep RBVF can be approximated easily and accurately. Moreover, deep RBVFs can represent any true value function owing to their support for universal function approximation. We extend the standard DQN algorithm to continuous control by endowing the agent with a deep RBVF. We show that the resultant agent, called RBF-DQN, significantly outperforms value-function-only baselines, and is competitive with state-of-the-art actor-critic algorithms.