SLIM_GSGP: The Non-bloating Geometric Semantic Genetic Programming

Geometric semantic genetic programming (GSGP) is a successful variant of genetic programming (GP), able to induce a unimodal error surface for all supervised learning problems. However, a limitation of GSGP is its tendency to generate offspring larger than their parents, resulting in continually growing program sizes. This leads to the creation of models that are often too complex for human comprehension. This paper presents a novel GSGP variant, the Semantic Learning algorithm with Inflate and deflate Mutations (SLIM_GSGP). SLIM_GSGP retains the essential theoretical characteristics of traditional GSGP, including the induction of a unimodal error surface and introduces a novel geometric semantic mutation, the deflate mutation, which generates smaller offspring than its parents. The study introduces four SLIM_GSGP variants and presents experimental results demonstrating that, across six symbolic regression test problems, SLIM_GSGP consistently evolves models with equal or superior performance on unseen data compared to traditional GSGP and standard GP. These SLIM_GSGP models are significantly smaller than those produced by traditional GSGP and are either smaller or of comparable size to standard GP models. Notably, the compactness of SLIM_GSGP models allows for human interpretation.