Model-Based Optimization

Abstract

While intuition and educated guesses can be sufficient for determining the qualitative, structural specification of a distributed robotic system, modeling is often necessary when it comes to finding the optimal parameters of the said specification. However, most classical optimization schemes are unable to deal with the combined effects of non-convexity, discontinuity, and stochasticity found in models of SMPs at low abstraction level. Even macrodeterministic models generated in a bottom-up fashion are in principle nonconvex, and may exhibit numerous local minima that are difficult to deal with. In such cases, one needs to recourse either to optimization meta-heuristics such as Genetic Algorithm (GA) or Particle Swarm Optimization (PSO) (and, more specifically, their noise-resistant variants [212]) or to systematic searches of the parameter space. Both approaches require underlying models that exhibit an excellent balance between computation cost and accuracy, as they involve numerous evaluations of candidate solutions. Optimization metaheuristics can deal with parameter spaces of high dimensionality, but they are often used as black box methods. Instead, systematic searches become difficult to use with more than three parameters, but they offer more insights into the global, qualitative behavior of the system, which is very important from a design perspective.