NK Model

The NK model is a mathematical model of fitness landscapes introduced by Stuart Kauffman and Simon Levin to study the ruggedness of the landscape as a function of two parameters: N (the number of genes or components in the system) and K (the number of epistatic interactions — the number of other genes that influence each gene's fitness contribution). When K=0, the landscape is smooth with a single peak; when K=N-1, the landscape is maximally rugged and uncorrelated — every local step is as likely to decrease fitness as increase it.

The NK model's central finding is that Evolution faces a fundamental tension between exploitability and expressibility: a low-K landscape is easy to climb but has low fitness peaks, while a high-K landscape has higher peaks but is nearly impossible to navigate by Natural Selection. The model predicts that biological genomes should evolve toward intermediate K values — a regime sometimes called the edge of chaos — where the landscape is rugged enough to harbour high-fitness solutions but smooth enough to be navigable.

This connects directly to Self-Organization: Kauffman argued that biological organisms are not merely products of selection but also of self-organizing attractors in gene regulatory networks. The landscape an organism evolves on is not fixed — it is itself co-constructed by the organism's developmental architecture, suggesting that Evolvability and Self-Organization are not independent phenomena but aspects of the same underlying dynamic.\n== NK Landscapes and the Dynamics of Search ==\n\nThe NK model is not merely a static description of fitness landscapes; it is a dynamical systems model of search. A genetic algorithm, an adaptive walk, or any evolutionary process traversing an NK landscape is a dynamical system whose state is the current genotype and whose dynamics are governed by the local gradient of the fitness function. The ruggedness parameter K determines the topology of this dynamical system: low-K landscapes have a single global attractor (easy convergence), while high-K landscapes fragment into many local attractors (trapping search).\n\nThis dynamical perspective reveals why the NK model matters beyond evolutionary biology. Any optimization algorithm — whether biological evolution, simulated annealing, or neural network training — operates on an implicitly defined landscape. The neural computation perspective treats synaptic weight space as a high-dimensional landscape shaped by the loss function, and training as a trajectory through that landscape. The NK model's insight — that landscape structure determines search dynamics — applies directly: deep neural networks face their own ruggedness problem, with local minima, saddle points, and flat regions that trap or slow gradient descent.\n\nThe connection to dynamical systems theory is precise. An NK landscape can be viewed as a potential function, and evolutionary search as a noisy gradient descent on that potential. The edge of chaos regime — intermediate K — corresponds to a dynamical regime where the system has enough structure to guide search but enough complexity to avoid premature convergence. This is the same trade-off that appears in reservoir computing, where the spectral radius of the reservoir determines whether dynamics are contractive (too simple) or chaotic (too unstable). The NK landscape is not merely a model of biological evolution. It is a general theory of how structure and search interact — and that generality is its deepest contribution.