Wiki

Evolutionary Algorithms

How evolutionary algorithms connect to game AI, evolutionary deep learning, prompt search, optimization, and agent systems.

Evolutionary algorithms are search methods for trying candidate solutions when the target can be scored but not directly derived. Micheal Lanham connects them to game AI and numerical optimization. He also connects them to evolutionary deep learning, prompt search, and modern AI agents (^[1]).

As an algorithm family, evolutionary algorithms center on fitness functions, mutation, and selection. They also cover architecture search, prompt search, and optimization tradeoffs. For the career and history bridge from game AI into modern agent workflows, use the related Game AI to LLM Agents page.

Evolutionary algorithms sit near machine learning, deep learning, and reinforcement learning. The shared structure is search under feedback. A team defines a fitness signal and generates candidate solutions. It keeps stronger candidates, then mutates or combines them until the result converges or the compute budget runs out (^[1]).

Search Mechanics

Genetic algorithms use a population of candidates, a fitness function, a mutation function, and sometimes a pairing function for combining parents. Fitter candidates reproduce, random mutations create new variants, and the search continues until convergence or resource limits (^[1]).

The practical tradeoff is compute. Evolutionary algorithms became popular across many applications because they explore many possible solutions with few hard constraints, but they’re computationally intensive. Around 2006 researchers treated them as a possible path to intelligence. Later deep learning frameworks became dominant (^[1]).

Game AI and Industry Optimization

Game-like environments come before any generic algorithm catalog. Early academic work built a game for testing children’s executive function. The same work then used simple neural networks and evolutionary algorithms to create test sequences and analyze player data (^[1]).

The method belongs to a simulated interaction. The team can try candidate tests, collect behavior, and compare outcomes.

The industrial example is more direct optimization. After oil-and-gas product work, evolutionary algorithms handled numerical analysis related to pipeline corrosion. They adapted faster and handled data efficiently. Their compute cost and weaker fit with the frameworks that boosted deep learning helped push them aside (^[1]).

A broader algorithms-and-optimization toolkit places evolutionary algorithms alongside graphs. For optimization, it names random sampling, gradient descent, and simulated annealing. It also names genetic algorithms for permutation problems (^[2]).

The combined picture is narrow but useful: evolutionary algorithms aren’t a replacement for mainstream ML. They’re search techniques for cases where teams can score candidates and afford repeated trials.

Evolutionary Deep Learning

The book “Evolutionary Deep Learning” by Micheal Lanham combines deep learning with evolutionary algorithms. The concrete uses include hyperparameter search and network architecture modification, especially for convolutional neural networks (^[1]). Evolutionary algorithms belong to model selection and architecture tuning rather than everyday supervised modeling.

Weight training is separate from design search: a CNN learns from data, while an evolutionary method searches over architecture variants or hyperparameters. These approaches can work well, but they’re computationally intensive (^[1]).

That makes evaluation part of the design problem. Teams need baselines and deployment constraints before spending compute on architecture or hyperparameter search. They also need a clear decision the model supports.

Prompt Search and Generative AI

The most modern example is prompt engineering, where evolutionary algorithms apply to prompts for LLMs and agents. The system generates prompt variants, scores the results, and evolves toward prompts that perform better (^[1]). This is promising because LLM behavior is complex and prompt variants can produce unexpected outputs.

Prompt search can be computationally expensive. One small example repo may take about a week (^[1]). It belongs with the broader prompt engineering and LLM evaluation workflows pages. Prompt evolution needs a scoring method, and the compute cost has to buy better behavior. For the lineage from game environments to LLM agent behavior, see Game AI to LLM Agents.

APIs and Tooling

EVOL is an API-design example. It’s an evolutionary algorithm library built to simplify genetic algorithms, which often become nested for loops. It used population and evolution objects with a functional API so the algorithm was easier to use (^[3]).

This is about open-source tool design, not a new theory of evolutionary search. It shows why API design matters for algorithm families that involve populations, scoring, mutation, and repeated generations. A clearer API can make the search easier to maintain even when the underlying method remains compute-heavy.

Optimization and Decision Systems

Decision optimization sets the boundary by separating prediction from action selection. A model’s .predict answers “what will happen.” A decision function answers “what should I do about it.” Airline pricing and fraud review thresholds are examples where business rules combine predicted probabilities with value or cost (^[4]).

Evolutionary search sits on that boundary. It can propose or tune candidates, but the team still needs a fitness function, simulator, or evaluation target. The same holds from the decision side. Teams simulate different decision rules and propagate outcomes over time. They rely on domain knowledge rather than a supervised model that tries to optimize the whole business objective (^[4]).

A cautionary engineering example comes from laser-system design: genetic algorithms first, then simple reinforcement learning around 2014. The search produced interesting ideas, but some were impractical to manufacture because the problem was poorly formulated outside the simulation (^[5]).

Search methods can optimize the wrong target when the simulator leaves out constraints, business objectives, or real-world costs. That laser-system example gives the decision-optimization boundary a physical engineering case.

Connection to Agent Systems

Collaborative agents can resemble evolutionary search when they generate candidate outputs, exchange feedback, and refine a result (^[1]). The comparison stays modest: a multi-agent system does not become an evolutionary algorithm only because agents collaborate. The useful connection is narrower. Agents generate candidates, exchange feedback, and refine outputs while teams still set stopping rules and compute limits.

Use Game AI to LLM Agents for Lanham’s transition from game AI into LLM agents. Use Multi-Agent Systems for coordination patterns among agents.