Paper Image

Reinforcement learning agents: Summarizing, comparing and interpreting their behavior

Published on:

17 January 2022

Primary Category:

Artificial Intelligence

Paper Authors:

Tom Bewley,

Jonathan Lawry,

Arthur Richards

Bullets

Key Details

Partitions state space into abstract states using data-driven axis splits

Aggregates transitions into time windows based on policy changes

Maximizes discrimination between policies using an information-theoretic objective

Enables intuitive visualization and analysis of agent dynamics over time

Agnostic to agent internals; applicable to any RL algorithm

AI generated summary

Reinforcement learning agents: Summarizing, comparing and interpreting their behavior

This paper introduces a novel technique to summarize and compare the behavioral dynamics of reinforcement learning agents over time. It works by partitioning the state space and aggregating transition data into abstract states and time windows. An information-theoretic objective is used to identify partitions that maximize the ability to discriminate between agent policies. The method produces interpretable graphical models of the agent's changing dynamics, enabling intuitive analysis of how and why its behavior evolves during learning.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up