Decent direction methods on the feasible region recognized by supervised learning metamodels to solve unstructured problems

Abstract

This paper treats with an unstructured optimization problem with three kinds of data: some feasible solutions, some infeasible solutions and an objective function. Then a decision support system (DSS) is developed to recognize the feasible region by using radial basis function (RBF) neural network, feed forward neural network, support vector machine (SVM) and decision tree algorithms. To optimize the problem, a penalty function is defined as a weighted combination of the objective function and the output of the feasible region recognition part. This function can be optimized using some gradient decent direction and quasi-Newton methods. The accuracy of feasible and infeasible solutions recognition of five types of optimization models are greater than 65% and in some cases 100%. In addition, the relative error of optimization part is 0.05%. Furthermore, toll pricing for Sadr Bridge is modeled by the aid of simulation results in AIMSUN where the accuracy of the proposed DSS is greater than 99%.

Subject Classifications:

• 68Uxx
• 90Cxx

Keywords:

• Nonlinear Programming
• Supervised Learning Algorithms
• Decent Direction Methods
• Decision Support System
• Toll Pricing
• Metamodel