Random search solutions of MINLP problems

WOS: 000071169800001The optimization models of chemical engineering processes are generally nonlinear and mixed integer nonlinear in structure and contain a lot of equality and inequality constraints and bounds on variables. In this work, two random search algorithms were developed based on the algorithms devised previously. The first algorithm uses a branch and bound (B&B) procedure along with random search procedure (RSP). Integer variables of MINLP (Mixed Integer Nonlinear Programming) problem were determined by B&B algorithm and NLP (NonLinear Programming) problems were solved by using a (RSP) at each feasible node. The second algorithm determines integer and continuous variables by the RSP. The algorithms include improvements as variable parameter dependent compression vectors for contraction, shifting strategies to move away from local optima and expansion trials. In the initial step of the algorithms, a coarse random search technique was used to provide three different feasible starting points. Then 75 global iterations were made to determine best starting vector. This increased the possibility of keeping away from a local optimum point. Finally 400 global iterations were made to determine the optimum value. The importance of the starting vector and the robustness of algorithms were searched statistically, by solving many complex examples of flowsheet optimization and process synthesis models, and mathematical problems using 50 different seeds in Borland C++ 3.1