A novel NeuroEvolutionary algorithm
Maryam Mahsal Khan, Gul Muhammad Khan · 2010
Cartesian Genetic Programming based Neuroevolutionary algorithm is proposed. It encodes the neural network attributes namely weights, topology and functions and then evolves them for best possible weight, topology and function. The architecture generated are both feedforward and recurrent. The proposed algorithm is applied on the standard benchmark control problem: balancing single and double pole at both markovian and non-markovian states. Results demonstrate that CGPANN has the potential to generate neural architecture and parameters in substantially fewer number of evaluations in comparison to earlier neuroevolutionary techniques. The power of CGPANN is its representation which leads to a thorough evolutionary search producing generalized networks. This opens new avenues of applying the proposed technique to any non-linear and dynamic problem.