{"paper":{"title":"Hierarchical Support Vector State Partitioning for Distilling Black Box Reinforcement Learning Policies","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Linear support vector machine splits distill black-box reinforcement learning policies into fewer interpretable subpolicies with higher returns.","cross_cats":["cs.HC"],"primary_cat":"cs.LG","authors_text":"Ann Now\\'e, Mehrdad Asadi, Senne Deproost","submitted_at":"2026-05-05T19:40:05Z","abstract_excerpt":"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%."},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"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%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That linear SVM splits on a distillation dataset of state-action pairs will reliably produce a compact hierarchical set of human-interpretable subpolicies that accurately mimic the original black-box policy behavior.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SVSP partitions distillation datasets with linear SVMs to create compact interpretable subpolicies, reporting +7.4% better mean return than VSP and +2.8% over TD3 while using 82.1% fewer subpolicies.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Linear support vector machine splits distill black-box reinforcement learning policies into fewer interpretable subpolicies with higher returns.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"70524e805331e949039010867d6f5c1df893ae0e5e454b60f61b66ebae80dded"},"source":{"id":"2605.04254","kind":"arxiv","version":2},"verdict":{"id":"77b359a9-51b5-40f3-bf61-4e7d9a6f4ca4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T16:35:48.801042Z","strongest_claim":"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%.","one_line_summary":"SVSP partitions distillation datasets with linear SVMs to create compact interpretable subpolicies, reporting +7.4% better mean return than VSP and +2.8% over TD3 while using 82.1% fewer subpolicies.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That linear SVM splits on a distillation dataset of state-action pairs will reliably produce a compact hierarchical set of human-interpretable subpolicies that accurately mimic the original black-box policy behavior.","pith_extraction_headline":"Linear support vector machine splits distill black-box reinforcement learning policies into fewer interpretable subpolicies with higher returns."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.04254/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T23:31:20.927215Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T14:40:39.029023Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"2e2e6e65576b471ba085e1a27885c9a510439f9f4665ed6b65a73c11ef9d28aa"},"references":{"count":8,"sample":[{"doi":"","year":2016,"title":"Ribeiro, M., Singh, S. & Guestrin, C. ” Why should i trust you?” Explaining the predictions of any classifier.Proceedings Of The 22nd ACM SIGKDD International Conference On Knowledge Discovery And Dat","work_id":"2f79d143-ed08-47ef-837c-658925c65f38","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Deproost, S., Steckelmacher, D. & Now ´e, A. Explainable RL Policies by Distilling to Locally- Specialized Linear Policies with V oronoi State Partitioning.ArXiv Preprint ArXiv:2511.13322. (2025)","work_id":"3abb18b5-9486-4b0e-872a-d3e6d5e3aff5","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Kohler, H., Delfosse, Q., Akrour, R., Kersting, K. & Preux, P. Interpretable and Editable Programmatic Tree Policies for Reinforcement Learning. (2024,10,28)","work_id":"32539da4-9262-42d2-8026-c48785d2dc4f","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Coppens, Y ., Efthymiadis, K., Lenaerts, T., Now ´e, A., Miller, T., Weber, R. & Magazzeni, D. Distilling deep reinforcement learning policies in soft decision trees.Proceedings Of The IJCAI 2019 Work","work_id":"30e9efa7-b445-41d6-9733-ed28e7b60373","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Blanco, V ., Jap ´on, A. & Puerto, J. Multiclass optimal classification trees with svm-splits.Machine Learning.112, 4905-4928 (2023)","work_id":"181576ba-35aa-443c-9821-01e10132aad0","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":8,"snapshot_sha256":"8cc31c485a3a1014f5a4decd887449b386bb05bfe8deee7f6cc127234da57e74","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}