Towards Automatic Generation of Evolution Rules for Model-Driven Optimisation

Abstract

Over recent years, optimisation and evolutionary search have seen substantial interest in the MDE research community. Many of these techniques require the specification of an optimisation problem to include a set of model transformations for deriving new solution candidates from existing ones. For some problems—for example, planning problems, where the domain only allows specific actions to be taken—this is an appropriate form of problem specification. However, for many optimisation problems there is no such domain constraint. In these cases providing the transformation rules over-specifies the problem. The choice of rules has a substantial impact on the efficiency of the search, and may even cause the search to get stuck in local optima. In this paper, we propose a new approach to specifying optimisation problems in an MDE context without the need to explicitly specify evolution rules. Instead, we demonstrate how these rules can be automatically generated from a problem description that consists of a meta-model for problems and candidate solutions, a list of meta-classes, instances of which describe potential solutions, a set of additional multiplicity constraints to be satisfied by candidate solutions, and a number of objective functions. We show that rules generated in this way lead to optimisation runs that are at least as efficient as those using hand-written rules.