Optimize - Nonlinear Programming Start
 Maximization of a convex objective function with linear constraints leads to difficulties because the continuous problem has local maxima that may not be global minima. We consider again the nonlinear program given earlier, but allow the variables to be continuous. The figure below shows the solution obtained with the Excel Solver starting from all 0's in the Values row. The Solver terminates immediately because this solution is a local maximum. Unfortuantely, this solution is not the global optimum. A common method to overcome this is to search for the global optimum by repeatedly using a nonlinear solver with different starting points. The Optimize add-in automates this process. Select Math Program from the Optimize menu and choose NLP Start from the options. A combinatorial form is constructed with a variable for each of the original NLP variables. This form is used to generate alternative starting solutions for the NLP. The add-in generates random discrete values between the lower and upper limits, places the results into the Value array of the NLP model and uses the Excel Solver to search for the local optimum. The Factor in cell W4 multiplies the discrete values in row 7 to obtain the starting solution in row 11. In the present case a factor of 1 is satisfactory. For other cases it may be necessary to choose a factor that makes the starting solutions representative of the range of solutions in the feasible region. To search for the global optimum, we choose Random as a search method and choose 20 as the number of solutions to generate. The figure below shows the combinatorial form for the last iteration. These values provide the starting solution for the Excel Solver. The solution obtained with the starting solution is the one below. This is the global optimum. The sorted solutions obtained from the 20 runs are shown below. Although the optimum was found by several starting solutions, several other local optimum were also discovered. For the NLP Start option, the solutions stored in this table are the local optimum solutions found rather than the discrete starting solutions. The Excel Solver seems to have a random element that can find different local optima from the same starting solution. There is no guarantee that the global optimum will be found with this procedure, but certainly the result is much more likely to be optimum that a single run taken from an arbitary start. The add-in was not successful in calling the Excel Solver with this add-in using Office X on a Mac. The method seems to work with Excel 2001 on a Mac or Excel XP using Windows. Other combinations have not been tested.

Operations Research Models and Methods
Internet
by Paul A. Jensen