On the control landscape topology
Résumé
Evolutionary algorithms are powerful tools to optimize parameters and structure of
control laws. However, these approaches are often very costly, or even prohibitive, for expensive
experiments due to long evaluation times and large population sizes. Reducing the learning
time, e.g. by decreasing the number of function evaluations, is a challenging problem as it often
requires additional knowledge on the objective function and assumptions. We address the need
to analyze these algorithms and guide their acceleration through examination of the search space
topology and the exploratory and exploitative nature of the genetic operators. We show how this
gives insights on the convergence and performance behavior of Genetic Programming Control for
the drag reduction of a car model (Li et al., 2016). Profiling machine learning algorithms, that
are very powerful but also more complex to analyze, aids the goal to increase their performance
and making them eventually feasible for a wide range of applications.