Global optimizers are those optimization methods which search the entire optimization space. These are best suited for large or complex parameter spaces, as well as well for situations where there is no clear starting point. The global optimizers in CST Studio Suite® are:
- Genetic Algorithm: Using an evolutionary approach to optimization, the Genetic Algorithm generates points in the parameter space and then refines them through multiple generations, with random parameter mutation. By selecting the “fittest” sets of parameters at each generation, the algorithm converges to a global optimum.
- Suitable for: Complex problem domains and models with many parameters.
- Particle Swarm Optimization: Another global optimizer, this algorithm treats points in parameter space as moving particles. At each iteration, the position of the particles changes, according not only to the best known position of each particle, but the best position of the entire swarm as well. Particle Swarm Optimization works well for models with many parameters.
- Suitable for: Models with many parameters.
- Covariance Matrix Adaptation Evolutionary Strategy: The CMA-ES is the most sophisticated of the global optimizers, and has relatively fast convergence for a global optimizer. With CMA-ES, the optimizer can “remember” previous iterations, and this history can be exploited to improve the performance of the algorithm while avoiding local optimums.
- Suitable for: General optimization, especially for complex problem domains.
In situations where the parameter space is small, or where the system is already near to the optimum, local optimizers can converge much faster, at the risk of finding only the local optimum.