A hybrid global and local search approach using prediction in simulation-based optimization

To obtain a more robust optimization technique, it is common to combine different search strategies in an attempt to compensate deficiencies of the individual algorithms. In particular, the characteristics of optimization problems from computational engineering raise the need for this kind of hybrid approach to increase the probability to at least approximate the global optimum of a given problem. Furthermore, the excessive run time of the solution process demands an approach to reduce the computational effort either by distributed computing techniques or by substituting expensive function evaluations. A hybrid approach is presented which substitutes actual simulations by neural network-based predictions of simulation results and combines a globally oriented search with a local search. To improve the quality of the prediction and the solution, the neural network used for prediction is adapted to the different steps of the algorithm. An example from computational engineering (groundwater management) is used to demonstrate the feasibility of the approach and to present first results.