Evolution of Optimal ANNs for Non-Linear Control Problems using Cartesian Genetic Programming.
Maryam Mahsal Khan, Gul Muhammad Khan, Julian Francis Miller · 2010
Abstract—A method for evolving artificial neural networks using Cartesian Genetic Programming (CGPANN) is proposed. The CGPANN technique encodes the neural network attributes namely weights, topology and functions and then evolves them. The performance of the algorithm is evaluated on the well known benchmark problem of double pole balancing, a non-linear control problem. The phenotype of CGP is transformed into ANN and tested under various conditions in the task envi-ronment. Results demonstrate that CGPANN has the ability to generalize neural architecture and parameters in substantially fewer number of evaluations in comparison to earlier neuro-evolutionary techniques. We have also tested the CGPANN for generalization with different initial states (not encountered during evolution) over a range of evolved genotypes and obtained good results.