Neural networks are trained for balancing 1 and 2 poles attached to a cart on a xed track. For one variant of the single pole system , only pole angle and cart position variables are supplied as inputs; the network must learn to compute velocities. All of the problems are solved using a xed architecture and using a new version of cellular encoding that evolves an application speciic architecture with real-valued weights. The learning times and generalization capabilities are compared for neural networks developed using both methods. After a post processing simpliication, topologies produced by cellular encoding were very simple and could be analyzed. Architectures with no hidden units were produced for the single pole and the two pole problem when velocity information is supplied as an input. Moreover, these linear solutions display good generalization. For all the control problems, cellular encoding can automatically generate architectures whose complexity and structure reeect the features of the problem to solve.
[1]
L. Darrell Whitley,et al.
Optimization Using Distributed Genetic Algorithms
,
1990,
PPSN.
[2]
Heinz Mühlenbein,et al.
The parallel genetic algorithm as function optimizer
,
1991,
Parallel Comput..
[3]
Alexis P. Wieland,et al.
Evolving Controls for Unstable Systems
,
1991
.
[4]
David R. Jefferson,et al.
Selection in Massively Parallel Genetic Algorithms
,
1991,
ICGA.
[5]
A. P. Wieland,et al.
Evolving neural network controllers for unstable systems
,
1991,
IJCNN-91-Seattle International Joint Conference on Neural Networks.
[6]
L. Darrell Whitley,et al.
Adding Learning to the Cellular Development of Neural Networks: Evolution and the Baldwin Effect
,
1993,
Evolutionary Computation.