How does OptQuest involves the "Performance Measures"?

Question

question

Sebastian Kemper asked Jul 6, '17 Jordan Johnson commented Jul 14, '17

How does OptQuest involves the "Performance Measures"?

Hello,

I have a question on OptQuest in addition with FlexSim.
OptQuest supports a "weighted" search strategy in FlexSim.
If there are Pareto optimality cases, how does OptQuest involves the "Performance Measures"?
If I set one measure to factor 2 and one measure to factor 1, how does OptQuest handle this setting?
I read the white paper for OptQuest, but at section 4.3. "Advanced Features" at "multiple objectives" there are no explicit definitions.
In the FlexSim Manual I also found nothing specific. I think that OptQuest calculates percental with those weight values when it finds a Pareto-Frontier, but thats just guessing. Do you can give me any explanations?

Regards Sebastian

Software Version:

FlexSim 17.1.1

optquest performance meas

Answer 1 · 2017-07-06T16:40:32Z

Jordan Johnson answered Jul 6, '17 Jordan Johnson commented Jul 14, '17

Unless performance measures are part of a constraint or an objective, OptQuest does not use those values as part of its search. Those values are recorded by FlexSim, but they don't influence the optimization.

Usually, however, performance measures are part of an objective function. This is where they will influence the OptQuest search. Often, an objective function might simply return a performance measure value. Then the user must tell OptQuest whether to minimize or maximize that value.

This brings me to your questions about multi-objective optimization. In FlexSim, there are two ways that you can use multiple objectives. The recommended way is to use the Pattern search option. The pattern search will show the pareto frontier, which shows the set of solutions that have the best trade-offs. For example, if you maximize revenue, and minimize cost, the pareto frontier will show you a range of solutions. The first solution will show the cheapest, lowest revenue option, completely favoring the cost objective. The last solution will show the most expensive, highest revenue option, completely favoring the revenue objective. For every other solution on the pareto frontier, you can know that there is no solution that is cheaper that give as good a revenue.

In weighted mode, OptQuest simply adds together all the objective functions (multiplied by their weights), and maximizes the result. You could manually do the same in a single objective, if you wanted. However, this search method is not recommended; correctly setting the weights is very difficult. It is much easier (and often much more informative) to have the optimizer show you the trade-off curve, and then allow you to choose the balance between objectives that you'd prefer.

· 3

Robin Brunner commented · Jul 07, 2017 at 01:42 PM

Thank you Jordan,

this helped us a lot! I tried to devise a formula with that information on how OptQuest creates this "super-objective". I put a picture of that formula in the attachment. This seems all to be clearer now for us. My question now is: What is OptQuest doing with the "Goal" and "Lower, Upper Bounds" in the weighted search mode?

Kind Regards Robin

0 ·

weighted-search-optquest.png (18.5 KiB)

Robin Brunner commented · Jul 14, 2017 at 12:14 PM

@jordan.johnson

Sorry for double posting, but can you or someone else confirm my formula? And what is OptQuest doing with the "Goal" and "Lower, Upper Bounds" in the weighted search mode? I can't really get on the point just by experimenting with the optimizer.

0 ·

Jordan Johnson ♦♦ Robin Brunner commented · Jul 14, 2017 at 03:34 PM

The part within the brackets is exactly true. I am not sure why the variable are factored in on the outside. Without that term, you have correctly shown the scoring function in weighted multi-objective search.

As far as the lower bound, upper bound, and goal, they are used to keep the optimizer on track more. One of the problems with weighted search is that one objective, even when weighted well, tends to dominate. The upper bound and lower bound are basically constraints that keep the optimizer from neglecting or favoring that objective too much. The goal provides additional scoring.

However, those are my best guesses. The documentation from OptQuest is scanty (on this topic, that is), and the OptQuest folks actually discourage using that feature. It is very difficult to get the optimizer to go the direction you want with weighted search. For multi-objective searches, the pattern mode is recommended. In fact, we are investigating the idea of just removing the weighted mode altogether in a future version of FlexSim.

0 ·

question