Evolutionary Agent-based Policy Analysis in Dynamic Environments Evolutionary Agent-based Policy Analysis in Dynamic Environments

Evolutionary algorithms (EAs) form a rich class of stochastic search methods that use the Darwinian principles of variation and selection to incrementally improve a set of candidate solutions (Eiben and Smith, 2003; Jong, 2006). Both principles can be implemented from a wide variety of components and operators, many with parameters that need to be tuned if the EA is to perform as intended. Tuning however requires effort, both in terms of time and computing facilities. When resources are limited we are interested to know how much tuning an EA requires to reach an intended performance, andwhich parameters aremost relevant in the sense that tuning them has the biggest impact on EA performance. Likewise, when designing an EA to simulate a real evolutionary process we would like to minimize the dependency of our simulation results on specific parameter values. In this case the amount of tuning required until the simulation behaves as intended indicates how plausible and realistic the simulation really is. To measure the amount of tuning that is needed in order to reach a given performance, we introduce the REVACmethod for Relevance Estimation and Value Calibration. While tuning the EA parameters in an automated and systematic manner, the method provides an information-theoretic measure on howmuch the tuning of each parameter contributes to overall EA performance. We evaluate its reliability and efficiency empirically on a number of test cases that reflect the typical properties of EA parameter spaces, as well as on evolutionary agent-based simulations. Finally we compare it to another tuning method, meta-GA. Parts of this chapter have been published in Nannen and Eiben (2007b,a); de Landgraaf et al. (2007).