RASP-Tuner

RASP-Tuner is intended to help when context is informative about the latent environment and can add harmful variance when context is irrelevant. Candidate generation blends the current parameter with a retrieved historical hint before scoring candidates. RASP-Tuner combines retrieval, historical parameter hints, soft prompts, a mixture-of-experts surrogate, and lower-confidence-bound candidate selection. RASP-Tuner targets online context-conditioned regret minimization in non-stationary black-box optimization. RASP-Tuner is positioned for non-stationary settings with recurring operational contexts that predict the active optimum.