Coevolutionary Computation

Coevolutionary computation is the subfield of evolutionary computation in which the fitness of an individual depends on the composition of the population itself. Unlike standard evolutionary algorithms, which assume a static fitness landscape, coevolutionary systems model scenarios where the quality of a solution is defined relative to other solutions — competitive games, predator-prey dynamics, host-parasite interactions, and adversarial machine learning.

The formal challenge of coevolutionary computation is the "mediocre stable state" problem: populations can converge to strategies that beat average opponents but fail against specialists. This has led to the development of fitness sharing, niche formation, and archive-based methods that maintain diversity by rewarding individuals for defeating previously unseen opponents. The Baldwin effect — the evolutionary advantage of learned behaviors that later become genetically assimilated — has been observed and exploited in coevolutionary systems as a mechanism for accelerating adaptation.

Coevolutionary computation is the algorithmic mirror of Red Queen dynamics in biology: the arms-race escalation that occurs when two species are each other's primary selective pressure. Whether coevolutionary algorithms can produce sustained open-ended innovation — rather than cycling through a finite set of strategies — remains one of the field's deepest open questions.