The tree experiment was designed to test out whether some characteristics of natural trees (branching, symmetry) could evolve as a consequence of the environment. The design of a tree in nature is a product of conflicting objectives: maximizing the exposure to light while keeping internal stability.
The experimental design for the tree has a narrow attachment base: Only three
knobs. This provides very little sustentation for cantilevering, so the structure
will have to be balanced to reach out. A ``light'' resource, coming from directions
up, left and right, has one value per column or row. Light is ``absorbed''
by the first brick it touches -- and the fitness points given are equal to
the distance from the absorption point to the x or y axis. The
highest fitness would be a structure reaching out to completely cover the left,
right and top borders (see fig. 2.29 and table 2.12).
There were no symmetry-oriented operators in our experiments, as could be, for example a ``reverse'' recombination operator that switched the orientation of a subpart. This means that symmetry is not encouraged by representational biases. Instead, the problem setup requires balancing the total weight of both sides. The tree did evolve, however, with a central symmetry with branches reaching out, by evolving the same type of solution separately on both sides.
The general layout of the evolved tree has several similarities with that of
a real tree: there is a (somewhat twisted) trunk, with branches that become
thinner as they reach out, and ``leaves'', bulky formations that maximize the
surface at the end of the branch.
|
The tree is, among all our experiments, the one that most clearly illustrates the emergence of nested levels of organization, key indicator of what we call complex organization