Evolution and search.

Several experiments here search for something: a body shape, a growth rule, a creature, a set of parameters. Most use evolutionary search, which borrows evolution’s loop without its biology: make variants of the current candidates, score each in the simulation, keep the better ones, repeat. It needs no formula for how a design’s score depends on its parameters, only a way to run it, which suits simulations of bodies in water. The first xenobots were designed this way.

Evolution strategy

A search that samples variants around the current best and moves toward the ones that scored higher.

Encoding

How a design is written down for the search: cell by cell (direct), or as a pattern-making function (a CPPN) that tends to give symmetric, repeated shapes.

Gradient

Where the simulation can be differentiated, the search can follow the exact slope instead, as when a neural cellular automaton is trained.

A search that rewards one goal finds one kind of answer. MAP-Elites (Mouret and Clune, 2015) keeps a map instead: the behaviors are divided into cells (how much a body turns, how far it pushes things), and each cell keeps the best design found that behaves that way. The result is a whole range of good designs, which can be tested on tasks the search never saw. This family is called quality-diversity search.

POET (Wang, Lehman, Clune and Stanley, 2019) evolves the problems along with the solutions. It keeps pairs of an environment and a design, makes new environments by changing old ones, keeps a new one only if it is neither too easy nor impossible for the current designs, and lets designs move to other environments when they do better there. The hope is a curriculum: stepping stones that lead to designs direct optimization can’t reach.

Evolution needs three things: variation, some individuals leaving more offspring than others, and offspring resembling their parents (heredity). Remove heredity and nothing accumulates. The Price equation splits any change in a population’s average trait exactly into a selection part (parents with the trait left more offspring) and a transmission part (offspring differ from their parents), which tells adaptation apart from drift and bias in copying.

Sources: Mouret & Clune, “Illuminating search spaces by mapping elites”; Wang et al., POET; Wang et al., Enhanced POET.

• Morphospace Atlas: a MAP-Elites map of soft bodies, tested on surprise tasks.
• Chip Archipelago: POET with bodies and microfluidic chips.
• Anthrobot Lenia and the Rejuvenation Field: searches over creatures and grown bodies.
• Biobot Tierra: evolution inside a simulation, with invented heredity and the Price equation.
• Reading the results: why searches are checked on fresh seeds.