Integrating Policy Summaries with Reward Decomposition for Explaining Reinforcement Learning Agents

Yael Septon, Tobias Huber, Elisabeth André, Ofra Amir

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

5 Scopus citations

Abstract

Explainable reinforcement learning methods can roughly be divided into local explanations that analyze specific decisions of the agents and global explanations that convey the general strategy of the agents. In this work, we study a novel combination of local and global explanations for reinforcement learning agents. Specifically, we combine reward decomposition, a local explanation method that exposes which components of the reward function influenced a specific decision, and HIGHLIGHTS, a global explanation method that shows a summary of the agent’s behavior in decisive states. Results from two user studies show significant benefits for both methods. We found that the local reward decomposition was more useful for identifying the agents’ priorities. However, when there was only a minor difference between the agents’ preferences, the global information provided by HIGHLIGHTS additionally improved participants’ understanding.

Original languageEnglish
Title of host publicationAdvances in Practical Applications of Agents, Multi-Agent Systems, and Cognitive Mimetics. The PAAMS Collection - 21st International Conference, PAAMS 2023, Proceedings
EditorsPhilippe Mathieu, Frank Dignum, Paulo Novais, Fernando De la Prieta
PublisherSpringer Science and Business Media Deutschland GmbH
Pages320-332
Number of pages13
ISBN (Print)9783031376153
DOIs
StatePublished - 1 Jan 2023
Externally publishedYes
Event21st International Conference on Practical Applications of Agents and Multi-Agent Systems, PAAMS 2023 - Guimaraes, Portugal
Duration: 12 Jul 202314 Jul 2023

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13955 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference21st International Conference on Practical Applications of Agents and Multi-Agent Systems, PAAMS 2023
Country/TerritoryPortugal
CityGuimaraes
Period12/07/2314/07/23

Keywords

  • Explainable AI
  • Neural Networks
  • Reinforcement Learning

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'Integrating Policy Summaries with Reward Decomposition for Explaining Reinforcement Learning Agents'. Together they form a unique fingerprint.

Cite this