On the generality of parameter tuning in evolutionary planning

Abstract

Divide-and-Evolve (DaE) is an original "memeticization" of Evolutionary Computation and Artificial Intelligence Planning. However, like any Evolutionary Algorithm, DaE has several parameters that need to be tuned, and the already excellent experimental results demonstrated by DaE on benchmarks from the International Planning Competition, at the level of those of standard AI planners, have been obtained with parameters that had been tuned once and for-all using the Racing method. This paper demonstrates that more specific parameter tuning (e.g. at the domain level or even at the instance level) can further improve DaE results, and discusses the trade-off between the gain in quality of the resulting plans and the overhead in terms of computational cost.

Instance-based parameter tuning for evolutionary AI planning

“…The controlled parameters of DaE are described in table 2. For a detailed description of these parameters, see [4]. The feature-set consists of 12 features.…”

“…Tuning only one instance has of course a sense if only that instance is to be solved. Parameters tuned for one instance however, may not be optimal for other instances, as [4] demonstrates. Furthermore, this paper also demonstrates that parameter tuning for several domains simultaneously is even more difficult, if at all possible.…”

“…Even generalizing parameters learned on one instance to another instance of the same domain (intra-domain generalization) might be problematic, as there are instances with very different complexity in the same domain. For example in [4] per-domain tuning was performed with the most difficult, largest instance, considered as a representative of the whole domain. However, it is clear from the results that these parameters were often suboptimal for the other instances.…”

“…same number of objects) may require different settings from domain to domain. The poor results with global tuning in [4] indicate that these are issues to be considered. One workaround for this generalization issue is to relax the constraint of finding a single universally optimal parameterset -that certainly does not exist -and to focus on learning a complex relation between instances and optimal parameters.…”

“…As described in the introduction, one could use for this sake REVAC, SPO, RACING or ParamILS. Because however, the parameter optimization will be done successively for several instances, the simple yet robust (1+1)-Covariance Matrix Adaptation Evolution Strategy [10], in short (1+1)-CMA-ES, was chosen, and used with its robust own default parameters, as advocated in [4]. CMA-ES is a maximum-likelihood based method estimating the so called natural gradient and improving efficiency this way over the conventional gradientbased methods.…”