A carregar...
Publicação
Exploring Crossover Operators in SLIM_GSGP: Towards More Compact and Generalizable Models: A multi-phase evaluation of novel crossover operators for efficient and interpretable evolutionary learning
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 9.53 MB | Adobe PDF |
Orientador(es)
Resumo(s)
Genetic Programming (GP) is renowned for its ability to evolve symbolic, humaninterpretable models. SLIM_GSGP (Semantic Learning with Inflate and deflate Mutations), a recent non-bloating variant of Geometric Semantic Genetic Programming
(GSGP) has shown strong potential in balancing accuracy with interpretability, while
maintaining GSGP’s property of inducing a unimodal error surface on any supervised
learning problem. Despite its promising performance, SLIM_GSGP remains underexplored in the literature. Most existing studies have focused on its mutation-based
design, leaving the role and potential impact of crossover, and genetic exchange, largely
unexamined. As a result, the contribution of crossover to model evolution in this context
is still not well understood. This thesis investigates the role of crossover in SLIM_GSGP
by proposing and evaluating a diverse set of 24 crossover operators, designed to explore both syntactic and semantic characteristics. Using a three-phase experimental
methodology, we assess these operators across predictive performance, model size, and
behavioral diversity. Our analysis includes Pareto-based ranking, complexity profiling,
and benchmarking against mutation-only SLIM_GSGP configurations. These findings
challenge the prevailing notion that crossover plays a minor role in GSGP, suggesting
that, when properly defined, crossover can still play a significant role in the evolutionary process. Several of the proposed operators frequently yield models that are not only
more compact, but also perform as well as or better than existing approaches in terms
of generalization to unseen data. By analyzing the mechanisms behind their success,
this work contributes to a deeper understanding of how crossover can be harnessed to
improve the effectiveness of SLIM_GSGP, transforming it from a traditionally sidelined
operator into a strategically crafted mechanism capable of fostering the evolution of
symbolic models that are not only accurate and compact but also interpretable.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced Analytics, specialization in Data Science
Palavras-chave
Genetic Programming SLIM_GSGP GSGP Crossover Operators Symbolic Regression Interpretability SDG 9 - Industry, innovation and infrastructure SDG 17 - Partnerships for the goals