Create a multi-start analytic

The Multi-Start Analytic helps you to analyze the results of optimizations from different start points when the multi-start option is enabled in the Optimization Starting Point page of the Run Settings dialog box.

Use the multi-start analytic to understand convergence behavior

1. Set up case structure

   

   Since understanding the convergence behavior involves modifying the model, it is sensible to make a child case to perform the work in. Create the multi-start analytic as a child of this case you want to investigate.

2. Configure run settings

   

   Use the Run Settings dialog to configure the number of starting points. The number to choose depends on how unstable your model is, as well as how long it takes to optimize a single case and the speed of your computer. Run as many start points as possible, but try to ensure that changes can be made and analyzed quickly.

3. Optimize model

   

   Optimize the model and use one of the charts in the Analysis section, such as the Economics Line Chart, to review the results.

4. Set up watch list

   

   To understand the differences between the solutions of each multi-start point create a watch list in the Stack View page. In the watch list choose items which are likely to be different between the different start points. Sensible items to choose include:

   • Process unit details

   • Stream routing details

   • Inventory details

     It is probably not useful to add purchases and sales. These are included in the other charts and can be reviewed there, and the differences in purchases and sales are driven by the bad structure (not the other way around) so each start point is likely to have different purchases and sales, but in themselves these will not highlight any poor structure.

5. Review watch items

   

   Review the watch items. Use the filtering functionality to find only results that are significantly different. These point to the areas that are driving the different economic solutions.

6. Identify potential problem structure

   

   Use your knowledge of the model to spot potential problems. For example in the above set the reformer is seen to behave differently in the different solutions, and in this particular model the reformer is a multi-base model which can sometimes cause poor convergence behavior. This gives a potential avenue of investigation.

7. Adapt or constrain structure

   

   In the parent case, make a change to the model to help reduce the impact of the structure which may be causing the convergence problems. For example, as the reformer above is a multi-base model, you could force the process model to only use one base (and therefore in a sense only have a single set of Base + Delta drivers). In this model this can be achieved by fixing the base interpolation parameter in each of the reformer modes.

   The changes to make depend on the modeling structure you have used.

8. Re-optimize and re-analyze

   

   After making the changes to restrict the structure re-optimize the model and re-analyze the results.