Evolutionary Learning

Evolutionary learning is an optimization and machine learning approach inspired by biological evolution: instead of a single solution, it maintains a whole population of candidate solutions that improves across successive generations. In each generation, a fitness function evaluates the quality of each candidate, selection favors the best, crossover combines information from two “parents” and mutation introduces random changes. Thus, generation after generation, the population tends to improve.

The evolutionary computation family

This idea groups together several historical currents. Genetic algorithms, popularized by John Holland in his 1975 book, represent solutions as bit strings. Evolution strategies (Rechenberg and Schwefel, in the 1960s) focused on the mutation of real-number vectors. And John Koza's genetic programming (1992) took the idea further: evolving the programs themselves.

How it applies to learning

In machine learning, these techniques serve to optimize hyperparameters and select features, evolving populations of configurations. Its most striking application is neuroevolution: evolving neural networks. The best-known method, NEAT (Stanley and Miikkulainen, 2002), is notable because it adjusts not only the network's weights but also its topology, which grows in complexity starting from a minimal structure.

Advantages over the gradient

The great asset of evolutionary learning is that it does not need gradients: it only requires being able to evaluate a solution's fitness. That makes it useful when the objective function is not differentiable, is noisy or is a “black box.” Moreover, its search is global—the population explores the space in parallel, with less risk of getting trapped in a local optimum—and it is easily parallelizable.

Its limits and its place today

In exchange, it is costly: it demands a great many evaluations of the fitness function and is usually less efficient than gradient-based methods, especially in high-dimensional problems. Tuning its parameters—population size, mutation and crossover rates—is an art. In 2026 it is enjoying renewed interest: it has been proposed as a scalable alternative to reinforcement learning and is combined with it in what is called Evolutionary Reinforcement Learning.