Evolutionary biologists have coined the term exaptation to describe a
commonplace occurrence in the evolution of life: a limb or organ evolves originally
for a particular function, but later on is recruited to fulfill a new function
[56]. When experimenting with our ``embedded solver'', that
encodes the simulation within the structure (section 2.2.7), we
stumbled upon the role that this change of use fulfills in the long bridge experiment.
|
|
The idea of ``encoding the simulation in the representation'' is to save time by not letting the simulator search through the space of all possible paths on the support network. Instead, only one path is allowed, as encoded in the representation.
Figure 2.24 shows the detail from one run of the long bridge experiment using this new variant simulator. In fig. 2.21 every joint has a virtual axis of rotation (stars) which links two bricks by two edges. But with the compact simulator, only one direction is allowed, and consequently each joint has only one edge, and loads are allowed to flow in just one direction.
Fig. 2.24 is a close-up of an evolved bridge which has no `base' part. The familiar cantilevered bar and counterbalance exist, but the base does not evolve. Why? Suppose a mutation or recombination creates a new brick at the point marked with X. A brick at such location would reinforce the base of the bridge, absorbing some of the weight of the cantilevered beam. This requires, however, a change of function of brick 2 above it. Currently, this brick rests over brick 1, balancing the beam. A base works in the opposite direction, transferring load from 1 to 2 to X then to 5 and finally to the ground. The rest of the counterbalance would act at brick 2 instead of 1. With the simulation-in-the-representation scheme, this is unlikely to happen, for it requires the chaining of four mutations. In the meantime, the long beam cannot be supported and thus the entire structure collapses. After a long beam evolves that relies on a counterbalance, the change of function does not happen.
We conclude that in this case, the use of a realistic simulator allowed a change of use (from counterbalance to support) that cannot happen with a more limited simulator. Encoding the part together with its use resulted in an impoverished relationship between agent and reality.
