Rethinking Evolutionary Optimization: Paying Attention to Intermediate Transitional Solutions for Certain Problems
Abstract
Population-based evolutionary computation (EC) algorithms have long drawn inspiration from natural evolution, yet one fundamental aspect is not often enforced in them: gradualism of populations over generations. Species do not leap to their optimal form in a single step; they transition through acceptable incremental states. In contrast, in most EC applications, the focus is often to judge how quickly and accurately the final optimal solution(s) are found and not how and with what gradualism these solutions were arrived at. This disconnect becomes critical in certain real-world problems where instead of just the final near-optimal solution, the focus is to discover how the currently practiced solution must be gradually transformed to the final solution. In a recent study, the so-called "innovation path" (IP) task and a novel bi-objective EC algorithm were proposed to find the ordered transformations. In this paper, we examine how the proposed IP-seeking algorithm utilizes the missing evolutionary principle in finding these transformations, and contrast it with the outcome of commonly-used EC algorithms. Through analytical discussions and illustrative scenarios, we show that only emphasizing converging to the optimal solution may not lead to the desired and acceptable transitional solutions. Our study highlights that if intermediate solutions are of importance, there is a need to rethink EC algorithms motivated by the gradual transitional aspect of natural evolution.
Authors: Ahmer Khan, Kalyanmoy Deb
Published in: Genetic and Evolutionary Computation Conference Companion (GECCO Companion) (2026)