Classification aided domain reduction for high dimensional optimization

Engineering design optimization often involves computationally expensive time consuming simulations. Although surrogate-based optimization has been used to alleviate the problem to some extent, surrogate models (like Kriging) struggle as the dimensionality of the problem increases to medium-scale. The enormity of the design space in higher dimensions (above ten) makes the search for optima challenging and time consuming. This paper proposes the use of probabilistic support vector machine classifiers to reduce the search space for optimization. The proposed technique transforms the optimization problem into a binary classification problem to differentiate between feasible (likely containing the optima) and infeasible (not likely containing the optima) regions. A model-driven sampling scheme selects batches of probably-feasible samples while reducing the search space. The result is a reduced subspace within which existing optimization algorithms can be used to find the optima. The technique is validated on analytical benchmark problems.