GPDL

Abstract

Defining custom problem types in genetic programming (GP) software systems is a tedious task that usually involves the implementation of custom classes and methods including framework-specific code. Users who want to solve a custom problem have to know the details of the targeted framework, for instance cloning semantics, and often have to write a lot of boilerplate code in order to implement the necessary functionality correctly. This can lead to frustration and hinders new developments and the application of GP to solve interesting problems. In this contribution we propose a framework-independent definition language for GP problems that can reduce the required effort and facilitate the integration of new problem types. We draw a parallel between the implementation of compilers for programming languages and the implementation of GP problems and reuse the well-established concept of attributed grammars with semantic actions to define computational symbols, semantics and structural constraints for GP. This goes beyond previous work in the area of context-free-grammar GP and grammatical evolution, because we also interweave the definition of symbol semantics and the target function with the definition of the grammar. This paper describes the proposed GP problem definition language (GPDL) and exemplary definitions of two popular benchmark problems using GPDL. We also describe a reference implementation of a GPDL compiler for HeuristicLab.