[attachment=9262]
Data Mining with Neural Networks and Genetic Algorithms
General Strategy
• Train Neural Network with representative data
• Use Genetic Algorithm to generate rule sets
• Evaluate performance of the rule sets
Training the Neural Network
• Only constraint on topology is that it must have the same number of inputs as categories (e.g., cap-shape, cap-surface, etc.) and the same number of outputs as classes (e.g., poisonous or edible)
• 3-layer perceptron appears to be model used for experiment with 7 input neurons, 6 hidden neurons, and 2 output neurons
• Backpropagation appears to be the weight modification rule
Using Genetic Algorithm to Generate Rule Sets
• Most important feature of a genetic algorithm is the fitness function – should:
– Achieve a high classification success rate
– Minimize relative entropy
– Reduce number of categories in rule
– Not reproduce other rules?
• Standard Goldberg model used for implementation
Evaluating Performance of Rule Sets
• Can use either fitness or majority function
• Majority function tends to classify more inputs correctly
Penalties and Benefits
• See5 beats rule sets under almost all conditions
• Under very high noise conditions, genetic extraction technique generalizes better
• Novelty of approach suggests lots of other possibilities