113
turning left and right, jumping and etc.) For insuring the control of the robot are very
important. The choice of templates is highly influenced by these factors.
9.3
Conclusion
This paper has presented a concept based on an evolutionary technique for the robot
locomotion learning. The technique proposed was a combination of both CNN and genetic
algorithms. The motivation of this combination can be justified by the high accuracy of the
CNN processors and their good computational speed as well. Further, the topology of CNN
is flexible for designing neuro-evolutive systems. The genetic algorithm was exploited for
the training process in order to determine the best genes according to the pre-defined
requirements (i.e. dada requirements) for the design process. Two types of robots were
considered (i.e. both structured and unstructured robots). For each of these types,
algorithms were developed to derive the appropriate chromosomes from which
corresponding templates were derived. The results in this paper have shown that
combining the cellular neural networks (CNN) technology with an evolution scheme like
genetic algorithm (GA) is very effective and suitable for learning the movement
/locomotion of different types of robots (e.g. high DOF robots, symmetrical, unsymmetrical
and defective robots). Due to the intrinsic characteristics of the CNN, this type of neural
network is very close to natural processors and therefore is efficient for building robot
controllers. During the training process, we found that the complexity of the environment
(e.g. rough, bumpy, and/or scaly surfaces) was a key factor influencing the results.
Basically, the technique developed in this paper provided interesting results with high
