Quantum feedback control is challenging to implement as a measurement on a quantum state only reveals partial information of the state. A feedback procedure can be developed based on a trusted model of the system dynamic, which is typically not available in practical applications. We aim to devise tractable methods to generate effective feedback procedures that do not depend on trusted models. As an application, we construct a reinforcement-learning algorithm to generate adaptive for quantum-enhanced phase estimation in the presence of arbitrary phase noise. Our algorithm exploits noise-resistant differential evolution and introducesan accept-reject criterion. Our robust method shows a path forward to realizing adaptive quantum metrology with unknown noise properties.