CrystalBox: Future-Based Explanations for Input-Driven Deep RL Systems

TL;DR

CrystalBox is a model-agnostic explainability framework for Deep Reinforcement Learning that generates future-based, high-fidelity explanations in input-driven environments, enhancing interpretability and practical utility.

Contribution

It introduces a novel algorithm for future-based explanations in DRL, leveraging decomposed returns and reward functions, applicable to both discrete and continuous control tasks.

Findings

Effective in adaptive bitrate streaming and congestion control.

Provides high-fidelity, contrastive, and guided explanations.

Demonstrates higher utility over prior explainability methods.

Abstract

We present CrystalBox, a novel, model-agnostic, posthoc explainability framework for Deep Reinforcement Learning (DRL) controllers in the large family of input-driven environments which includes computer systems. We combine the natural decomposability of reward functions in input-driven environments with the explanatory power of decomposed returns. We propose an efficient algorithm to generate future-based explanations across both discrete and continuous control environments. Using applications such as adaptive bitrate streaming and congestion control, we demonstrate CrystalBox's capability to generate high-fidelity explanations. We further illustrate its higher utility across three practical use cases: contrastive explanations, network observability, and guided reward design, as opposed to prior explainability techniques that identify salient features.