Towards an Evolutionary Design of Modular Robots for Industry - Foundations on Natural and Artificial Computation - page 55

3. Replacement: 10% of the individuals are eliminated from the population and
replaced by variations of the best 5% robots and by variations of random
robots. These variations are obtained through a symmetry mutation oper-
ator. This operator selects a branch of a robot chromosome and replicates
it in order to achieve a symmetrical branch. Rotational and reflective sym-
metry are currently implemented. If it is impossible to apply the symmetry
operator a random robot is generated.
4. Steps 2 and 3 are repeated until a stop criterion is met.
The mutation operators defined above perform a growth phase (add node),
a morphology adaptation phase (shake modules), a control adaptation phase
(shake control) and a pruning phase (delete nodes), being the order of these
phase the backbone of the proposed constructive approach. The default config-
uration implies applying each operator once per cycle, although the best results
were obtained applying the three structural operators in a higher degree than
the shake control operator, reinforcing the idea that morphology and control
must evolve in different time scales. The algorithm does not include a crossover
operator because we have found that, with the basic tree representation of the
modular systems we are using, crossing two individuals is basically a random
search procedure.
Regarding the calculation of fitness, it is carried out during the mutation
stage as each operator is applied. Thus, when step 3 is reached, all the individ-
uals have already been evaluated. Fig. 2 displays a diagram of the evolutionary
framework used to evolve and evaluate the modular robots. It is made up of the
evolutionary strategy detailed above, which is included in the Java Evolutionary
Algorithm Framework (JEAF) [11] developed within our group and accessible
Fig. 2. Evolutionary design system