8.6 Comparison of Local Optimizers and Evolutionary Computation

Although evolutionary computation is a well established optimization technique today, its application to TCAD analysis has been limited. Reasons are certainly the need for lots of computational resources and the requirements outlined in Chapter 9. While most research in evolutionary computation has been done on relatively cheap to evaluate goal functions, the optimization of semiconductor devices has to cope with a relatively limited number of evaluations.

The most important difference from the usual practice of evolutionary algorithm optimizers is that runs are usually finished before a common termination condition like ``95% of the population are identical'' is fulfilled.

In the following we discuss the advantages and disadvantages of local optimizers and evolutionary algorithm optimizers and show why the combination of both is worthwhile. An optimization run with an evolutionary algorithm optimizer usually yields a set, or population, of nearly optimal solutions. The best of these is used as the starting point for a run with a gradient based optimizer, thus bringing together global and local optimization methods. Other combinations are also possible, for example starting populations can be constructed manually; other states than the best in a population may yield better final results because they lie closer to the global optimum; the configuration of an optimizer can be changed and the computation restarted with the latest population or starting point.