Parameterized reinforcement learning for optical system optimization

Abstract Engineering a physical system to feature designated characteristics states an inverse design problem, which is often determined by several discrete and continuous parameters. If such must particular behavior, the mentioned combination of both, continuous, parameters results in challenging optimization problem that requires extensive search for optimal design. However, if corresponding can be reformulated as parameterized Markov decision process, reinforcement learning (RL) provides heuristic framework solve it. In this work, we use multi-layer thin films example aforementioned problems consider three parameters: Each film layer’s dielectric material (discrete) thickness (continuous), well total number layers (discrete). While recent methods merely determine thicknesses and—less commonly—the layers’ materials, our approach optimizes stacked well. summary, further develop Q-learning variant thereby outperform human experts current approaches like needle-point or naive RL. For purpose, propose exponentially transformed reward signal eases policy enables constrained optimization. Moreover, learned Q-values contain information about optical properties films, allows us interpretation what-if analysis thus explainability.