Home // COGNITIVE 2023, The Fifteenth International Conference on Advanced Cognitive Technologies and Applications // View article
Authors:
Torben Logemann
Eric MSP Veith
Keywords: reinforcement learning; explainable AI; equivalent transformation; neural network; decision tree; compression
Abstract:
Learning systems have achieved remarkable success. Agents trained using Deep Reinforcement Learning (RL) (DRL) methods, e.g., promise real resilience. However, no guarantees can yet be provided for the learned black-box models. For operators of Critical National Infrastructures (CNIs), this is a necessity as no responsibility can be assumed for an unknown and unvalidatable control system. Intrinsically secure learning algorithms and approximate, post-hoc interpretable models exist, but they lack either learning performance or explainability. To optimize this trade-off, this paper presents the NN2EQCDT algorithm, which equivalently transforms a Feed-Forward Deep Neural Network (DNN) (FF-DNN)-based policy into a compressed Decision Tree (DT). The compression is achieved by dynamically checking the satisfiability of the paths during construction, removing checks that are not needed further, and considering invariants. For a small policy model, NN2EQCDT was observed to drastically compress the DT, making it possible to accurately trace action regions to their observation regions in a plotted DT and visualization.
Pages: 94 to 100
Copyright: Copyright (c) IARIA, 2023
Publication date: June 26, 2023
Published in: conference
ISSN: 2308-4197
ISBN: 978-1-68558-046-9
Location: Nice, France
Dates: from June 26, 2023 to June 30, 2023