Constraint-Handling with Support Vector Decoders

A comparably new application for support vector machines is their use for meta-modeling the feasible region in constrained optimization problems. Applications have already been developed to optimization problems from the smart grid domain. Still, the problem of a standardized integration of such models into (evolutionary) optimization algorithms was as yet unsolved. We present a new decoder approach that constructs a mapping from the unit hyper cube to the feasible region from the learned support vector model. Thus, constrained problems are transferred into unconstrained ones by space mapping for easier search. We present result from artificial test cases as well as simulation results from smart grid use cases for real power planning scenarios.