Policy composition in reinforcement learning via multi-objective policy optimization

TL;DR

This paper introduces a multi-objective policy optimization framework that leverages pre-existing teacher policies to accelerate reinforcement learning and improve task performance, especially when shaping rewards are absent.

Contribution

It presents a novel method for composing and extending teacher policies within a multi-objective reinforcement learning setting, including a mechanism for agents to select teachers dynamically.

Findings

Teacher policies accelerate learning in continuous domains.

Agents can compose and extend teacher policies to solve complex tasks.

Dynamic teacher selection improves task reward in humanoid control.

Abstract

We enable reinforcement learning agents to learn successful behavior policies by utilizing relevant pre-existing teacher policies. The teacher policies are introduced as objectives, in addition to the task objective, in a multi-objective policy optimization setting. Using the Multi-Objective Maximum a Posteriori Policy Optimization algorithm (Abdolmaleki et al. 2020), we show that teacher policies can help speed up learning, particularly in the absence of shaping rewards. In two domains with continuous observation and action spaces, our agents successfully compose teacher policies in sequence and in parallel, and are also able to further extend the policies of the teachers in order to solve the task. Depending on the specified combination of task and teacher(s), teacher(s) may naturally act to limit the final performance of an agent. The extent to which agents are required to adhere to teacher policies are determined by hyperparameters which determine both the effect of teachers on learning speed and the eventual performance of the agent on the task. In the humanoid domain (Tassa et al. 2018), we also equip agents with the ability to control the selection of teachers. With this ability, agents are able to meaningfully compose from the teacher policies to achieve a superior task reward on the walk task than in cases without access to the teacher policies. We show the resemblance of composed task policies with the corresponding teacher policies through videos.