Abstract
We propose a scheme leveraging reinforcement learning to engineer controlfields for generating non-classical states. It is exemplified by theapplication to prepare spin-squeezed states for an open collective spin modelwhere a linear control field is designed to govern the dynamics. Thereinforcement learning agent determines the temporal sequence of controlpulses, commencing from a coherent spin state in an environment characterizedby dissipation and dephasing. Compared to the constant control scenario, thisapproach provides various control sequences maintaining collective spinsqueezing and entanglement. It is observed that denser application of thecontrol pulses enhances the performanceof the outcomes. However, there is aminor enhancement in the performance by adding control actions. The proposedstrategy demonstrates increased effectiveness for larger systems. Thermalexcitations of the reservoir are detrimental to the control outcomes. Feasibleexperiments are suggested to implement the control proposal. The extension tocontinuous control problems and another quantum system are discussed. Thereplaceability of the reinforcement learning module is also emphasized. Thisresearch paves the way for its application in manipulating other quantumsystems.