Concepts with a long memory

Evolutionary AI.

Evolutionary computation is not a relic of early artificial intelligence. It is a family of techniques for search, adaptation and design that continues to complement modern machine learning.

Evolutionary computation

Instead of improving a single solution by following a direct gradient, an evolutionary algorithm maintains candidate solutions and applies processes analogous to variation and selection. The approach is useful when a problem is discontinuous, noisy, multiobjective or difficult to differentiate.

A simplified cycle

  1. Create a population of candidate solutions.
  2. Evaluate each candidate against one or more objectives.
  3. Select promising candidates.
  4. Generate variation through mutation, recombination or other operators.
  5. Repeat, preserving useful diversity while improving performance.

Genetic algorithms

Genetic algorithms encode candidate solutions and evolve them through selection, crossover and mutation. They have been applied to scheduling, engineering design, control, games and many other optimisation problems represented at CEC2004.

Swarm intelligence

Swarm methods model collective behaviour. Particle Swarm Optimisation lets candidate solutions move through a search space using their own experience and information shared by neighbours. Ant Colony Optimisation uses reinforced paths inspired by pheromone trails. A surviving CEC2004 special-session page shows both methods already being treated as a rapidly growing research area.

Multiobjective optimisation

Real decisions rarely have a single objective. A faster design may cost more; an accurate model may consume more energy. Multiobjective evolutionary algorithms search for diverse trade-offs, helping decision makers understand the shape of the problem rather than hiding it behind one score.

Artificial life and emergent systems

Artificial life uses computational models to study adaptation, emergence and lifelike behaviour. It overlaps with agent-based systems, evolutionary robotics and open-ended exploration — areas where the process of discovery can matter as much as a single winning solution.

Where it meets modern AI

Contemporary AI is often associated with neural networks trained by gradient descent. Evolutionary methods offer a different toolbox and can complement learning systems.

  • Neural architecture search
  • Neuroevolution
  • Agent behaviour
  • Quality-diversity
  • Black-box optimisation
  • Robotics
  • Automated design
  • Hyperparameter search

The useful historical lesson is not that one family of methods replaces another. It is that intelligence can be engineered through several kinds of search, learning and adaptation.