Multimodal Adaptive Graph Evolution
Camilo De La Torre, Kévin Cortacero, Sylvain Cussat‐Blanc, Dennis G. Wilson · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
The problem of program synthesis involves automatically finding a function based on evaluation criteria, like matching input-output pairs. While Cartesian Genetic Programming (CGP) has excelled in various function synthesis tasks, it has primarily been limited to single data types, hindering its applicability to diverse data. Mixed-Type CGP, proposed in 2012, aimed to address this limitation but faced challenges due to search space limitations and complexity in building function libraries. In this study, we introduce Multimodal Adaptive Graph Evolution (MAGE), a generalized CGP extension that integrates functions of different data types by grouping them accordingly and imposing mutation constraints based on type. Through comparisons with standard CGP and Mixed-Type CGP on Program Synthesis Benchmark and image classification tasks, we demonstrate that MAGE's representation and mutation constraints facilitate the search for multimodal functions.