Deep Reinforcement Learning (DRL) achieves state-of-the-art for controller generation. However, it relies on a monolith neural policy to perform, reducing transparency and user trust. We distill a DRL policy into a set of human-readable models by partitioning the state space in regions. Using a hierarchy of linear Support Vector Machines (SVM) we show a close approximation of the original policy using simple linear functions
Deproost, S, Nowe, A & Asadi, M 2026, 'Distilling Linear Control Policies Using Linear Hierarchical SVMs', The 5th International Conference on Hybrid Human-Artificial Intelligence, Brussel, Belgium, 8/07/26 - 10/07/26.
Deproost, S., Nowe, A., & Asadi, M. (2026). Distilling Linear Control Policies Using Linear Hierarchical SVMs. Poster session presented at The 5th International Conference on Hybrid Human-Artificial Intelligence, Brussel, Belgium.
@conference{01ffdd0309674c07bd8ce947da9fdb07,
title = "Distilling Linear Control Policies Using Linear Hierarchical SVMs",
abstract = "Deep Reinforcement Learning (DRL) achieves state-of-the-art for controller generation. However, it relies on a monolith neural policy to perform, reducing transparency and user trust. We distill a DRL policy into a set of human-readable models by partitioning the state space in regions. Using a hierarchy of linear Support Vector Machines (SVM) we show a close approximation of the original policy using simple linear functions",
author = "Senne Deproost and Ann Nowe and Mehrdad Asadi",
year = "2026",
month = jul,
day = "6",
language = "English",
note = "The 5th International Conference on Hybrid Human-Artificial Intelligence, HHAI'26 ; Conference date: 08-07-2026 Through 10-07-2026",
url = "https://hhai-conference.org/2026/",
}