ALTERNATIVE EXAM [3 hours, 4 points]: A feedforward neural network architecture is given by a sequence of numbers of nodes in successive layers, e.g., (9, 7, 11, 12, 5) describes a network for a 9-input 5-output problem, with 3 hidden layers containing 7, 11, and 12 nodes. Implement a GA whose goal is to discover the “best” network architecture for a given problem, minimizing
(1-accuracy) . (# hidden layers) . max(1, max(# nodes in any hidden layer – # input nodes)).
For example, if accuracy is 90% for a (9,7,11,12,5) network, then this evaluates to ((1-0.9)(3)(12-9)) = 0.9, better than a (9,11,14,5) network with 80% accuracy, since ((1-0.8)(2)(14-9))=2.0.
1. Use PSO or ES to train any network to obtain the accuracy.
2. Report the results obtained on a non-trivial problem with high input dimensionality, i.e., where good results are not obtained easily using a simple network.
3. Summarize your results, discussing the usefulness of this approach.