Author Topic: Umberto and Optimization  (Read 4858 times)

Offline ignizio

  • New Member
  • *
  • Posts: 1
    • View Profile
Umberto and Optimization
« on: November 18, 2009, 17:26 »
I\'m in the process of evaluating the features of various simulation software packages. I\'ve used numerous packages in the past (ranging from GERT to SLAM to ARENA to EXTEND, etc.) but not Umberto.

I do find the documents and papers concerning Umberto to be of considerable interest and it certainly would appear that it is a powerful tool for analysis. But I\'m really concerned about the claims that Umberto can be used for OPTIMIZATION.

For example, there are several papers that claim that Umberto may be used to OPTIMIZE semiconductor and solar cell factories. Having spent 7 years as a staff scientist for a major semiconductor firm I find that claim somewhat difficult to accept.

OPTIMIZATION (i.e., true Optimization) of a system guarantees the development of a solution that is the very best possible solution subject to the system\'s constraints. Simulation (of any kind), however, simply evaluates a finite number of alternative configurations. From these one may select the best solution ---- but certainly not the OPTIMAL solution,

For example,in a 300mm semiconductor wafer fabrication facility there are more than 500 process steps, 80 or more functional areas (known as tool sets), and a infinite or near infinite number of possible solutions (e.g., tool selection, dispatch methods, supply types, run rules, etc., etc.).

To guarantee a true OPTIMAL soluton via simulation would require the evaluation of the infinite or near infinite combinations. Since each simulation of such a factory takes on the order of a day or more, we have to resort to evaluating a finite number of configurations (e.g., perhaps 50 or less). The probability of finding the true OPTIMAL configuration is, thus, zero.

So, my question is, how can Umberto guarantee the true OPTIMAL solution? Does it employ a (true) OPTIMZATION algorithm?

The only true OPTIMIZATION algorithm [i.e., that may be used for determining the capacity of such complex situations as found in semiconductor fabs] that I am aware of is found in Chapter 13 of the book: Ignizio, J.P., OPTIMIZING FACTORY PERFORMANCE, published by McGraw-Hill in 2009).

Is Umberto capable of incorporating such an OPTIMIZATION algorithm? This would require access to either a Linear or Nonlinear Optimization software package (e.g., such as CPLEX)?

Offline hendrik.lambrecht

  • New Member
  • *
  • Posts: 2
    • View Profile
Re: Umberto and Optimization
« Reply #1 on: November 27, 2009, 11:02 »
Dear Mr. Ignizio,

in principle, your observation is absolutely correct. “True optimization”, as you call it, i.e. an optimization procedure that can guarantee to find the global optimum (like e.g. the well known simplex algorithm for an LP), is not yet possible with Umberto. The term “process optimization” employed by many Umberto users, mainly from the industry (e.g., refers generally to analyzing, comparing and finally choosing the best out of different simulated scenarios.

However, there are current efforts to make “true” optimization available for Umberto-users in the medium term. A research project (KOMSA) on this topic ( has been finished recently. The main outcome is a prototypical optimization module for Umberto. As Umberto-models generally contain algorithmic function specifications that cannot be transferred into a completely analytical form, i.e. a mathematical program, its design is based on the simulation based optimization approach.  .

You might argue that simulation-based optimization has the drawback of not providing certainty for finding global optima as well. Nevertheless it strongly supports model analysis in practical applications and has become sort of a standard for many simulation software packages in recent years. As you mentioned the simulation software ARENA you may take the OptQuest-module as an example.

The implemented architecture of the KOMSA optimization module provides interfaces for integrating any algorithm or solver that supports direct (or at least derivative free) optimization. Mainly for the purposes of testing our approach, four different algorithms of that sort have been implemented in KOMSA: a genetic algorithm, particle swarm optimization, a strategy based on both the Nelder-Mead and the Complex-strategy of Box and finally a complete enumeration. Furthermore there are some techniques for model analysis, such as the calculation of directional derivatives for the decision variables. The module has also been tested in practice with first promising results (e.g.

Best regards,
Hendrik Lambrecht
Pforzheim University