Learning Global Optimization Through a Support Vector Machine Based Adaptive Multistart Strategy

We propose a global optimization algorithm called GOSAM (Global Optimization using Support vector regression based Adaptive Multistart) that applies statistical machine learning techniques, viz. Support Vector Regression (SVR) to adaptively direct iterative search in large-scale global optimization. At each iteration, GOSAM builds a training set of the objective function’s local minima discovered till the current iteration, and applies SVR to construct a regressor that learns the structure of the local minima. In the next iteration the search for the local minimum is started from the minimum of this regressor. The idea is that the regressor for local minima will generalize well to the local minima not obtained so far in the search, and hence its minimum would be a ‘crude approximation’ to the global minimum. This approximation improves over time, leading the search towards regions that yield better local minima and eventually the global minimum. Simulation results on well known benchmark problems show that GOSAM requires significantly fewer function evaluations to reach the global optimum, in comparison with methods like Particle Swarm optimization and Genetic Algorithms. GOSAM proves to be relatively more efficient as the number of design variables (dimension) increases. GOSAM does not require explicit knowledge of the objective function, and also does not assume any specific properties. We also discuss some real world applications of GOSAM involving constrained and design optimization problems.