Evolving High-Level Imperative Program Trees with Genetic Programming

Genetic Programming (GP) is a technique which uses an evolutionary metaphor to automatically generate computer programs. Although GP proclaims to evolve computer programs, historically it has been used to produce code which more closely resembles mathematical formulae than the well structured programs that modern programmers aim to produce. The objective of this thesis is to explore the use of GP in generating high-level imperative programs and to present some novel techniques to progress this aim.

A novel set of extensions to Montana’s Strongly Typed Genetic Programming system are presented that provide a mechanism for constraining the structure of program trees. It is demonstrated that these constraints are sufficient to evolve programs with a naturally imperative structure and to support the use of many common high-level imperative language constructs such as loops. Further simple algorithm modifications are made to support additional constructs, such as variable declarations that create new limited-scope variables. Six non-trivial problems, including sorting and the general even parity problem, are used to experimentally compare the performance of the systems and configurations proposed.

Software metrics are widely used in the software engineering process for many purposes, but are largely unused in GP. A detailed analysis of evolved programs is presented using seven different metrics, including cyclomatic complexity and Halstead’s program effort. The relationship between these metrics and a program’s fitness and evaluation time is explored. It is discovered that these metrics are poorly suited for application to improve GP performance, but other potential uses are proposed.