Black-BoxOptimizationwithLocalGenerativeSurrogates SupplementaryMaterial ASurrogatesImplementationDetails

To address this issue we rewrote one of the base layers that were used in FFJORD library. Latin Hypercube sampling window of size = 0.2 was used for the "Rosenbrock", "SubmanifoldRosenbrock", "Nonlinear Submanifold Hump", and"Neural NetworkWeightsOptimization" problems, =0.5wasusedfor"ThreeHump"problem. Muons are bent by the magnetic field and simultaneously experience stochastic scattering asthey pass through the magnet which causes random variations in their trajectories. The magnet is constructed from six trapezoidal shapes with gaps each of which is described by seven parameters, aspresented inFigure 8. Thus, forthis taskψ R42. Wefoundthatlearningrate0.1 is optimal for all the toy problems under consideration.