Exploring modular reuse methods and adaptive operator selection in genetic programming for solving boolean problems

Abstract

Standard genetic programming struggles to scale to complex problems due to rapidly growing search spaces. This thesis investigates modular reuse strategies to improve the scalability of tree-based genetic programming for Boolean function synthesis. Four approaches are studied: random module reuse, special initialization with transferred modules, hierarchical module reuse, and a bandit-based adaptive operator selection method. These approaches are evaluated across four benchmark tasks-parity, majority-on, adder, and multiplexer-which exhibit diverse structural properties. Results show that modular reuse significantly improves performance over standard GP, particularly for compositional tasks. Hierarchical reuse performs well on parity and multiplexer problems, while the bandit-based method demonstrates strong adaptability when suitable building blocks are available. However, limitations arise in tasks such as adder due to changing input-output dimensionality and in multi-output settings where reward assignment becomes unreliable. The study also investigated the importance of parsimony pressure in controlling solution bloat while using modular reuse methods.