On Search Trajectory Networks for Graph Genetic Programming
Camilo De La Torre, Sylvain Cussat‐Blanc, Dennis G. Wilson, Yuri Cossich Lavinas · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Cartesian Genetic Programming (CGP) allows for the optimization of interpretable function representations. However, comprehending the vast and combinatorially complex search space inherent to CGP remains challenging, particularly because multiple genotypes may correspond to identical functions. This paper studies the application of Search Trajectory Networks (STNs) to understand the search dynamics of CGP, specifically for symbolic regression tasks. Using STNs, we analyze the behavior of evolutionary search processes and uncover distinct phenomena, such as the presence of "portal" minima---critical junctures that facilitate sudden, beneficial shifts in the search trajectory, akin to findings in linear genetic programming. Our findings illustrate that while genetic interpretations are complex and often ambiguous, a functional analysis using STNs offers clear and actionable insights into the CGP search.