Support-boundary and relative-reliability diagnostics for hybrid memory–neural reinforcement learning
APPLIED SOFT COMPUTING JOURNAL, cilt.204, ss.116350, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 204
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.asoc.2026.116350
- Dergi Adı: APPLIED SOFT COMPUTING JOURNAL
- Derginin Tarandığı İndeksler: Applied Science & Technology Source, Scopus, Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC
- Sayfa Sayıları: ss.116350
- Marmara Üniversitesi Adresli: Evet
Özet
Hybrid reinforcement-learning controllers must distinguish supported memory use from unsupported neural extrapolation. The study develops an auditable support-boundary diagnostic protocol and evaluates relative-reliability arbitration. In a 30-seed application-shift test, k-nearest-neighbor and Gaussian-affinity support improve return AUC over exact count gating by 0.422 and 0.766, but remain below tabular Q-learning. A locked recurring-state shift favors fuzzy relative-reliability gating over count gating; across three independently locked generators, that same primary fuzzy-versus-count contrast yields two null results and one negative result. The co-primary same-input crisp-versus-fuzzy contrast is generator-specific and favors crisp in the localized generator. Sensorized Crazyflie/PyBullet software-in-the-loop testing exposes a failure boundary: exact support collapses, tolerant kNN yields partial coverage without waypoint success, and learned controllers remain at zero success. A separately locked supplemental, non-confirmatory benchmark evaluates SAC, CrossQ, and TQC on HalfCheetah-v5 and Walker2d-v5 (30 agents; 3600 final episodes). No controller-return contrast survives Holm correction. Observation-delay support differences are negative in all six algorithm–environment cells, but no P2 support contrast survives Holm correction; a 0.75 actuation-gain shift has small, mixed support effects. Thus support is representation- and shift-specific, not calibrated confidence or a generic detector of degradation. The contribution is a reproducible diagnostic framework preserving positive, null, and negative evidence rather than a universal controller or deployment claim.