Galapagos: Single-Objective Search

Drop the Galapagos component onto the canvas (Params tab, under Util). It has exactly two inputs and that simplicity is the whole point. Genome takes the sliders it is allowed to change - wire one or several number sliders into it, and each becomes a gene. Fitness takes a single number: the output of your definition that measures how good the current design is.

Double-click the component to open the Galapagos editor - a separate window with a Solvers tab (settings and the two engines) and a Record tab (the live view of the run). In the options you set whether you are maximising or minimising the fitness, and you tune the engine. Then you press Start, and Galapagos begins writing values into your sliders, letting the definition re-solve, reading the resulting fitness, and repeating - hundreds or thousands of times a minute for a light definition.

The mental model matters: Galapagos does not understand your geometry at all. It sees only genes going in and a fitness coming out. It is turning knobs blindfolded, guided solely by whether the number improved. Everything clever - the geometry, the meaning - lives in your definition between the genome and the fitness.

This blindness is worth sitting with, because it explains most beginner frustration. Galapagos will not tell you your fitness is nonsense, that a slider does nothing, or that the winning shape is hideous - it cannot see any of that. It is a knob-turner with a single readout. The intelligence of the whole exercise is front-loaded into how you build the definition and, above all, how you define the number that comes out. Get those right and Galapagos is remarkably powerful; get them wrong and it will confidently optimize your mistake.

A good Galapagos setup is mostly good wiring. Choose your genes deliberately. Every slider you add multiplies the search space, so include only the parameters that genuinely matter, and set their ranges sensibly - too wide and the solver wastes runs on absurd designs; too narrow and it cannot reach the good region. Integer sliders (counts) and floating sliders (angles, depths) both work as genes.

Collapse everything you care about into one fitness number. With a single objective you get exactly one wire into Fitness, so if you care about more than one thing you must combine them - typically a weighted sum, for example 0.7 * shading_score + 0.3 * view_score, or you fold constraints in as penalties: subtract a big number when the design breaks a rule so the solver learns to avoid it. This is the moment to be honest, because whatever you put in that wire is exactly what Galapagos will pursue.

Before running, exercise the definition by hand: drag each gene slider across its range and confirm the fitness responds sensibly and the definition never errors. A run that hits a null geometry halfway through is a wasted afternoon. When the manual behaviour looks right, you are ready to let the solver take the wheel. A minute spent driving the sliders by hand routinely saves an hour of a long run quietly chasing a broken or meaningless fitness.

Galapagos ships with two search engines, and knowing when to pick each is worth real time.

The Evolutionary Solver - the default, and the one David Rutten describes at length in his classic Evolutionary Principles applied to Problem Solving essay - keeps a population of genomes. Each generation it evaluates them, keeps the fittest, and breeds new genomes by combining parents and applying small random mutations. Because many individuals are spread across the landscape, it explores broadly and is comparatively good at avoiding local optima. Its key options are population size (more = broader search, slower), and the maintain/inbreeding sliders that trade exploration against exploitation.

Simulated Annealing works differently: it holds a single point and wanders, accepting improvements always and worse moves sometimes - often early, when the 'temperature' is high - so it can climb out of a local dip. As it cools, it accepts fewer downhill moves and settles into a peak. It can be faster on smoother problems and where you want to refine near a known-good region. Neither is universally better. A sound habit: start with the Evolutionary Solver to explore, note the promising region, then optionally anneal to refine. And run it more than once - if several runs agree, trust the result more.

Do not just stare at the best number. In the Record tab, watch the convergence: fitness should climb steeply at first, then flatten. A curve still rising fast means keep going; a curve that has been flat for many generations means the solver has converged and extra runtime buys little - stop it. If it plateaus almost immediately at a poor value, suspect a landscape with a dominant local trap, a fitness that barely responds to your genes, or ranges that are too tight.

Then comes the essential discipline: look at the winning geometry, not only its score. A single-objective solver is a literal-minded optimizer, and it will happily over-fit - exploit any loophole in your fitness to inflate the number in ways you never intended. Maximise daylight and it may dissolve the whole facade into glass, brilliant on your metric and a solar-gain disaster you did not measure. Minimise structural depth and it may find a razor-thin section that scores well and buckles in reality. When the optimum looks wrong, the fitness is wrong. This is not a flaw in Galapagos; it is the nature of pointing a relentless searcher at one number. Add the missing term, re-penalise the loophole, and run again - the loop between reading the result and fixing the objective is where the real design thinking lives.

A few practical habits turn Galapagos from a toy into a dependable instrument. The single biggest lever is definition speed: Galapagos re-evaluates your entire definition once for every candidate, thousands of times, so anything slow inside it - a dense mesh, a heavy Boolean, a full daylight simulation - is multiplied enormously. Before optimizing, strip the definition to the essentials, coarsen any analysis meshes for the search and refine only the winner, and switch off unrelated components so they are not re-solving alongside. A run that would take hours often drops to minutes after this trim.

On the engine options, resist the urge to over-tune. Sensible defaults and a larger population generally beat clever fiddling, because population is what buys exploration and protects you from local traps. Give the run enough generations to actually flatten - stopping too early is the commonest mistake - but do not sit watching a converged curve crawl for hours of no gain.

Most importantly, run it more than once. Evolutionary search is stochastic: two runs from different random seeds can settle on different peaks, especially on a rugged landscape. If several independent runs agree, you can trust the result; if they scatter widely, that itself is telling you the landscape is rugged or your fitness barely responds to the genes. Treating a single run as gospel is how people get burned - reproducibility is your cheapest and most honest quality check, and it costs nothing but a little patience.