A0C: Alpha Zero in Continuous Action Space

TL;DR

This paper extends Alpha Zero's tree search and deep learning framework to handle continuous action spaces, demonstrating preliminary success in robotic control tasks and paving the way for broader real-world applications.

Contribution

It introduces theoretical modifications to Alpha Zero for continuous actions and provides initial experimental validation on a control task.

Findings

Feasibility of Alpha Zero extension to continuous actions

Preliminary success on Pendulum swing-up task

First step towards real-world continuous domain applications

Abstract

A core novelty of Alpha Zero is the interleaving of tree search and deep learning, which has proven very successful in board games like Chess, Shogi and Go. These games have a discrete action space. However, many real-world reinforcement learning domains have continuous action spaces, for example in robotic control, navigation and self-driving cars. This paper presents the necessary theoretical extensions of Alpha Zero to deal with continuous action space. We also provide some preliminary experiments on the Pendulum swing-up task, empirically showing the feasibility of our approach. Thereby, this work provides a first step towards the application of iterated search and learning in domains with a continuous action space.