Summarising and Comparing Agent Dynamics with Contrastive Spatiotemporal Abstraction

TL;DR

This paper presents a data-driven, model-agnostic method for summarizing and contrasting the dynamics of evolving systems like reinforcement learning agents, using spatiotemporal aggregation based on information divergence.

Contribution

It introduces a novel technique for generating human-interpretable summaries of agent dynamics, applicable to continuous state spaces and enhancing interpretability.

Findings

Effective summarization of deep RL learning histories

Complementary to existing interpretability methods

Applicable to continuous state spaces

Abstract

We introduce a data-driven, model-agnostic technique for generating a human-interpretable summary of the salient points of contrast within an evolving dynamical system, such as the learning process of a control agent. It involves the aggregation of transition data along both spatial and temporal dimensions according to an information-theoretic divergence measure. A practical algorithm is outlined for continuous state spaces, and deployed to summarise the learning histories of deep reinforcement learning agents with the aid of graphical and textual communication methods. We expect our method to be complementary to existing techniques in the realm of agent interpretability.