(Emergent Morphology)

7 Evolution

A population evolves towards improved average fitness, where an animat's fitness is proportional to the distance traveled from the location where an animat starts at the beginning of the animation. Direction of travel does not matter. The secondary fitness terms concerning head activity can contribute to the overall fitness. To determine the fitness of an animat, the animation is run for 1000 time steps, in which the animat is essentially dropped into an infinite, flat world. The animat has a limited life span in which to gather distance.

A population of 100 animats is used in each experiment. The population is not updated in parallel as in the standard GA but rather in increments of one individual at a time as shown in the steps below:

Initialize: Set all gene values to be random, then animate and evaluate each animat to determine all initial fitness values

Iterate: Cycle through the following sequence 1000 times:
* Each genotype is ranked by fitness within the population
* Two genotypes are selected randomly with chances of being chosen proportional to ranking, and they mate via crossover.
* They produce one offspring which is subject to some mutation
* The new offspring is animated and evaluated for fitness
* Its genotype replaces the lowest ranking genotype

Limiting reproduction to one mating at a time causes slower convergence than updating in parallel but it has the benefit of inhibiting premature convergence, and it guarantees that the most fit genotypes at any given time are never lost. Mutation rate is set at inverse proportion to population size.

8 Results

(go to beginning of document)