Genetic Algorithms

In 1975, computer scientist John Holland published a book with a quiet but radical idea: what if you could solve hard optimization problems the way life solves hard environments — not by reasoning from first principles, but by evolving a population of candidates toward better and better answers?

The recipe is almost biological. Start with a random population of candidate solutions. Score each one with a fitness function. Let the fitter candidates reproduce more often. When two candidates reproduce, swap chunks of their representations (crossover). Occasionally flip a random bit (mutation). Repeat for many generations.

No single step requires deep understanding of the problem. There is no guarantee the best solution is ever found. And yet, generation after generation, the population climbs toward surprisingly good answers in spaces that would take the age of the universe to search exhaustively.

That combination — powerful in practice, impossible to guarantee in theory — is what makes genetic algorithms both fascinating and tricky to reason about.

The demo below runs a classic phrase-evolution experiment. The target is a short string. A population of random strings evolves toward it one generation at a time through selection, crossover, and mutation.

<div class="ga-wrap">
  <div class="target-row">
    <label>{{lbl_target}}</label>
    <input id="target" type="text" value="HELLO WORLD" maxlength="20" spellcheck="false"/>
  </div>
  <div class="params-row">
    <label>{{lbl_popsize}} <input id="popsize" type="number" min="10" max="200" value="60"/></label>
    <label>{{lbl_mutrate}} <input id="mutrate" type="number" min="1" max="50" value="4"/>%</label>
  </div>
  <div class="btns">
    <button id="btnStart" type="button">{{btn_start}}</button>
    <button id="btnStep"  type="button">{{btn_step}}</button>
    <button id="btnReset" type="button" class="ghost">{{btn_reset}}</button>
  </div>
  <div class="stats">
    <span>{{lbl_generation}} <b id="genNum">0</b></span>
    <span>{{lbl_best_fit}} <b id="bestFit">—</b></span>
  </div>
  <div class="best-label">{{lbl_best_ind}}</div>
  <div id="bestStr" class="best-str">—</div>
  <canvas id="chart" width="480" height="120"></canvas>
  <div id="statusMsg" class="status"></div>
</div>

* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; padding: .5rem; }
.ga-wrap { max-width: 480px; }
.target-row, .params-row { display: flex; align-items: center; gap: .6rem; flex-wrap: wrap; margin-bottom: .45rem; font-size: .88rem; }
.target-row label, .params-row label { color: #555; }
input[type="text"] { font: 600 1rem ui-monospace, monospace; padding: .28rem .5rem; border: 1px solid #bbb; border-radius: 6px; width: 12rem; text-transform: uppercase; }
input[type="number"] { width: 4.5rem; padding: .28rem .4rem; border: 1px solid #bbb; border-radius: 6px; font-size: .88rem; }
.btns { display: flex; gap: .4rem; flex-wrap: wrap; margin: .3rem 0 .6rem; }
button { font: 600 13px system-ui, sans-serif; padding: .4rem .85rem; border: 1px solid #1d3557; background: #1d3557; color: #fff; border-radius: 7px; cursor: pointer; }
button.ghost { background: #fff; color: #1d3557; }
.stats { font-size: .88rem; color: #444; display: flex; gap: 1.2rem; margin-bottom: .3rem; }
.stats b { color: #1d3557; }
.best-label { font-size: .8rem; color: #777; margin-bottom: .15rem; }
.best-str { font: 700 1.25rem ui-monospace, monospace; letter-spacing: .08em; min-height: 1.8rem; margin-bottom: .5rem; }
.best-str .match { color: #0a7d33; }
.best-str .miss  { color: #c92f3c; }
canvas { display: block; border: 1px solid #dde3ea; border-radius: 8px; background: #f7f9fb; margin-bottom: .4rem; }
.status { font-size: .9rem; font-weight: 600; min-height: 1.4em; color: #0a7d33; }

// Code not found

Notice how the population improves fast at first, then slows as the remaining wrong characters become harder to fix. The algorithm has no idea what the target looks like — it only knows the fitness score. This is the same dynamic that plays out in every real GA: rapid early improvement, then a race against diminishing returns and local optima.

Genetic algorithms are heuristics — they search without a proof of optimality.

• Status: heuristic (no optimality guarantee). There is no general theorem that a GA will find the global optimum of a given problem. For specific problem families, convergence proofs exist, but they come with conditions that are hard to verify in practice.
• Why they still work. The schema theorem (Holland, 1975) gave an informal argument: short, high-fitness patterns ("schemata") propagate exponentially faster than random sampling, effectively parallelizing the search across many promising subspaces at once.
• The traps. GAs can converge prematurely when diversity is lost too fast, get stuck in local optima, or waste generations on regions of the search space that look fit but are not. Tuning population size, mutation rate, and crossover operator is as much art as science.
• Theoretical limits. The No Free Lunch theorem (Wolpert & Macready, 1997) shows that no single algorithm — including genetic algorithms — outperforms random search on average across all possible problems. Every gain on one class of problems is paid back on another. GAs win by exploiting the structure of real-world problems, not by magic.
• Connection to hard problems. GAs are often applied to NP-hard problems like the Traveling Salesman or job scheduling — problems where exact algorithms are too slow. The GA does not solve them exactly, but often finds solutions within a few percent of optimal, which is enough for engineering.

This is the honest summary: genetic algorithms are a powerful, widely applicable search strategy with strong empirical track records and weak theoretical guarantees. Understanding both sides is what separates practitioners who trust them wisely from those who trust them blindly.

Any domain with a large search space and a measurable quality score is a natural home for genetic algorithms:

• Engineering design: NASA's Space Technology 5 antenna (2006) was designed by a GA — its strange, bent shape emerged from evolution, not an engineer's blueprint, yet it outperformed hand-designed alternatives.
• Scheduling and logistics: airline crew rosters, factory job scheduling, and vehicle routing all benefit from GA-based solvers when the problem is too large for exact methods.
• Drug discovery and protein design: GAs explore the chemical space of molecules, evolving candidates with target binding properties faster than purely random screening.
• Neural architecture search (NAS): evolutionary methods search the space of neural network architectures, discovering designs that rival or beat hand-crafted ones — a key tool in the modern AI toolbox.
• Game AI and strategy: GAs have evolved strategies for classic games, trading agents, and robot locomotion gaits, finding behaviors that human designers would not have predicted.
• Parameter tuning: even when not solving the core problem, GAs tune the hyperparameters of other algorithms, including machine learning models.

The common thread: wherever the search space is vast, the objective is computable but hard to optimize analytically, and approximate answers are acceptable, genetic algorithms provide a robust, parallelizable toolkit.

Genetic algorithms carry a humbling message: sometimes the most effective way to find a good answer is not to think harder but to breed better candidates. No deep insight required — just a fitness function, a population, and enough generations.

The price is honesty: GAs cannot prove they found the best solution, they can get stuck, and the No Free Lunch theorem reminds us there is no universally superior algorithm. But in the messy, high-dimensional spaces of engineering, biology, and AI, that is often a trade worth making.

Evolution built the human brain in roughly four billion years of iterations. Genetic algorithms try to capture the same logic in milliseconds. They do not always succeed — but when they do, the results can look as alien and elegant as a bent antenna optimized for space.