A genetic algorithm is a search and optimization technique inspired by the principles of natural selection, in which a population of candidate solutions is iteratively evolved through operations analogous to mutation, crossover (recombination), and selection to improve fitness toward a defined objective over successive generations. Genetic algorithms belong to the broader family of evolutionary computation methods within Artificial intelligence and Machine learning.
Mechanism
A genetic algorithm begins with a randomly generated population of candidate solutions, each typically encoded as a string or vector analogous to a chromosome. Each candidate is evaluated using a fitness function that scores how well it solves the target problem. Higher-scoring candidates are more likely to be selected as "parents," whose encodings are combined through crossover to produce offspring, with random mutations occasionally introduced to maintain diversity and avoid premature convergence on a suboptimal solution. This cycle of evaluation, selection, and recombination repeats across many generations, with the population as a whole tending to improve in average fitness over time, though not guaranteed to find a global optimum.
History
The field's mathematical foundations were formalized by John Holland, whose 1975 book "Adaptation in Natural and Artificial Systems" introduced genetic algorithms as a general framework for adaptive search, building on earlier evolutionary computing experiments from the 1950s and 1960s. Holland's students and collaborators, including David Goldberg, extended the framework's theoretical grounding and popularized practical applications through the 1980s and 1990s.
Applications
Genetic algorithms have been applied to scheduling and routing problems, engineering design optimization, including antenna and aerodynamic shapes evaluated by NASA and others, automated program synthesis under the related field of genetic programming, and hyperparameter search for machine learning systems. They are particularly favored for problems with large, complex, non-differentiable search spaces where gradient-based methods are unavailable or ineffective, since genetic algorithms require only the ability to evaluate a candidate's fitness, not to compute a derivative of the objective.
Neuroevolution
A notable application area, neuroevolution, uses evolutionary methods to design or train Neural network architectures and weights, sometimes combined with Reinforcement learning for control and game-playing tasks. Neuroevolution has been explored as an alternative or complement to backpropagation-trained networks in Robotics and Embodied AI research, where the reward signal is sparse or the network topology itself, not just its weights, is a design variable.
Limitations and modern relevance
Genetic algorithms scale poorly to the very high-dimensional parameter spaces of modern deep networks compared to Gradient descent-based optimization such as Backpropagation, and largely receded from mainstream research as backpropagation-trained Deep learning architectures came to dominate after the early 2010s. They remain actively used, however, in optimization domains outside standard supervised training, in neuroevolution niches, and as a conceptual reference point for research on open-ended and evolutionary approaches to generating diverse, novel solutions rather than optimizing a single fixed objective.