Most advances on the Evolutionary Algorithm optimisation of Neural Network are on recurrent neural network using the NEAT optimisation method. For feed forward network, most of the optimisation are merely on the Weights and the bias selection which is generally known as conventional Neuroevolution. In this research work, a simultaneous feature reduction, network pruning and weight/biases selection is presented using fitness function design which penalizes selection of large feature sets. The fitness function also considers feature and the neuron reduction in the hidden layer. The results were demonstrated using two sets of data sets which are the cancer datasets and Thyroid datasets. Results showed backpropagation gradient descent error weights/biased optimisations performed slightly better at classification of the two datasets with lower misclassification rate and error. However, features and hidden neurons were reduced with the simultaneous feature /neurons switching using Genetic Algorithm. The number of features were reduced from 21 to 4 (Thyroid dataset) and 9 to 3 (cancer dataset) with only 1 hidden neuron in the processing layer for both network structures for the respective datasets. This research work will present the chromosome representation and the fitness function design.
[1]
Yoshua Bengio,et al.
Gradient Flow in Recurrent Nets: the Difficulty of Learning Long-Term Dependencies
,
2001
.
[2]
Jürgen Schmidhuber,et al.
Long Short-Term Memory
,
1997,
Neural Computation.
[3]
Geoffrey E. Hinton,et al.
Learning representations by back-propagating errors
,
1986,
Nature.
[4]
Risto Miikkulainen,et al.
Evolving Neural Networks through Augmenting Topologies
,
2002,
Evolutionary Computation.
[5]
Kenneth O. Stanley and Bobby D. Bryant and Risto Miikkulainen,et al.
Real-Time Evolution in the NERO Video Game (Winner of CIG 2005 Best Paper Award)
,
2005,
CIG.
[6]
Georgios C. Anagnostopoulos,et al.
Interactively evolved modular neural networks for game agent control
,
2008,
2008 IEEE Symposium On Computational Intelligence and Games.