aiqbench takeMachine Learning

SVSP introduces a novel approach to distilling black-box reinforcement learning policies into compact, human-interpretable subpolicies using linear support vector machine splits. The method demonstrates significant efficiency gains, improving mean return by 7.4% over prior partitioning techniques while reducing the required number of subpolicies by over 80%. This work advances the field by enabling more flexible distillation where decision boundaries and surrogate models can be optimized within a margin of the original policy behavior.

Source coveringarXiv

Hierarchical Support Vector State Partitioning for Distilling Black Box Reinforcement Learning Policies

arXivJune 10, 20262 min read

arXiv:2605.04254v3 Announce Type: replace Abstract: We introduce State Vector Space Partitioning (SVSP), a novel method to mimic a black box reinforcement learning policy using a set of human-interpretable subpolicies. By partitioning a distillation dataset of state action pairs with linear support vector machine splits, SVSP constructs a compact and structured representation of the original policy. Our method improves mean return by +7.4% over previous critic driven state partitioning attempts such as Voronoi State Partitioning (VSP) and +2.8% over the original TD3 policy, while reducing the number of required subpolicies against VSP by 82.1%. Our results pave the path towards a more flexible form of distillation where both the decision boundary and surrogate models can be chosen within a margin of the original black box behavior.

Abstract:We introduce State Vector Space Partitioning (SVSP), a novel method to mimic a black box reinforcement learning policy using a set of human-interpretable subpolicies. By partitioning a distillation dataset of state action pairs with linear support vector machine splits, SVSP constructs a compact and structured representation of the original policy. Our method improves mean return by +7.4% over previous critic driven state partitioning attempts such as Voronoi State Partitioning (VSP) and +2.8% over the original TD3 policy, while reducing the number of required subpolicies against VSP by 82.1%. Our results pave the path towards a more flexible form of distillation where both the decision boundary and surrogate models can be chosen within a margin of the original black box behavior.

Machine Learning (cs.LG); Human-Computer Interaction (cs.HC)

arXiv:2605.04254 [cs.LG]

(or arXiv:2605.04254v3 [cs.LG] for this version)

Full text under the publisher's open licence · 134 words

Read at arXiv ↗Ask about this article