Mining Patterns from Genetic Improvement Experiments

When conducting genetic improvement experiments, a large amount of individuals (≈ population size * generations) is created and evaluated. The corresponding experiments contain valuable data concerning the fitness of individuals for the defined criteria, such as run-time performance, memory use or robustness. This publication presents an approach to utilize this information in order to identify recurring context independent patterns in abstract syntax trees (ASTs). These patterns can be applied for restricting the search space (in the form of anti-patterns) or for grafting operators in the population. Future work includes an evaluation of this approach, as well as extending it with wildcards and class hierarchies for larger and more generalized patterns.