{"id":"arxiv/2610.18524","url":"https://arxiv.org/abs/2610.18524","tags":["math.OC"],"score":0,"title":"Hamiltonian Primal-Dual Policy Reinforcement Learning for Derivative-Free Model-Based Control","domain":"math.OC","source":"arxiv","authors":["K. Rabiei","Z. Rui","X. Lu","W. Hao"],"reviews":0,"subject":"math.OC","summary":"Finite-horizon optimal control becomes difficult when a dynamical model can be simulated reliably, but its derivatives are unavailable, discontinuous, or unreliable. Hamiltonian Primal-Dual Policy Optimization (HPDPO) is introduced as a derivative-free, model-based reinforcement learning framework for this setting. The control is represented as a time-indexed finite-horizon policy. Candidate actions are evaluated by deterministic suffix rollouts of the supplied transition model, while a regularized Hamiltonian dual formulation supplies structured future-cost information through…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-13","reactions":{}}