Abstract
In several real-world engineering and industrial optimization problems, we only have historical data collected from physical experiments, numerical simulations, or production operations for performing optimization, and further, no expensive fitness evaluations are available during the optimization process. This class of optimization problems are called offline data-driven optimization. As no new data is available as offline data points, it is of utmost importance to build good quality surrogates from the initial data and harness useful information from unlabelled data to enhance the performance of the surrogates. This paper proposes a novel approach to combine the exploitation of local surrogate and exploration of global surrogate, guiding the optimization process to reach the optimum with the limited amount of available data. Furthermore, we follow a new method to choose reliable individuals for updating both global and local surrogates. Moreover, to show the high competitiveness of our algorithm over other state-of-the-approaches, we evaluate our model on problems up to 100 dimensions.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have