#mainTextBox Genetic optimization is a powerful tool for enhancing the efficiency and effectiveness of a neural network. NeuroSolutions provides the capability to genetically optimize virtually any component parameter(s) in a neural network. This tutorial will demonstrate the use of genetic optimization to improve the classification accuracy of a two hidden-layer MLP.
#subtitleTextBox MLP Without Genetic Optimization
#mainTextBox Below is shown a two hidden-layer (10 and 5 PEs respectively) MLP that attempts to detect a class composed of disjoint clusters in 3-dimensional space. The dataset is complicated by the presence of two channels of noise. Run the MLP and observe its inability to converge on an optimal classification. This inaccuracy stems primarily from the MLP's inability to discern the valid inputs from the noise inputs. The learning curve of the network and the classification performance of the network (both as a confusion matrix and graph of % correct classification) are provided below.
#subtitleTextBox Genetic Optimization - Adding the Genetic Control
#mainTextBox Genetic optimization can fine-tune network parameters so that network performance is greatly enhanced. To incorporate genetic optimization in any network, simply add the Genetic control component (as highlighted below). The Genetic control applies a genetic algorithm (GA) to any network parameters that are specified. Also through the Genetic control, GA parameters such as mutation probability, crossover type and probability, and selection type can be modified. Browse the Inspector below to investigate these modifiable parameters.
#subtitleTextBox Genetic Optimization - Selecting Component Parameters to Optimize
#mainTextBox Once the Genetic control has been added, it is necessary to specify which network parameters are to be optimized. The Inspector for each component now shows a 'Genetic' tab that provides a list of all optimizable component parameters. A check next to any of these will add the parameter to the GA list. A TanhAxon has been selected below; click on the 'Genetic' tab in the Inspector window and peruse the axon parameters that can be optimized. Note that for each selected parameter, mutation type and other intrinsic parameter values can be specified.
#mainTextBox Genetic optimization can even be used to select the most favorable input channels for network performance. In the File Inspector below, click on the 'Customize' button. A window is displayed that allows you to select input channels, (via the checkboxes labeled 'GA') which are to be genetically optimized. By selecting an input channel(s), GA will determine if that particular channel(s) is truly beneficial in the training of the network.
#subtitleTextBox MLP With Genetic Optimization
#mainTextBox The network below has already been enhanced using genetic optimization (this was done in order to reduce the time of this demo; genetic optimization can be a lengthy process depending on the speed of the machine). Right-click on one of the hidden axons, select the "Properties" menu item, then select the "Genetic" tab to inspect the parameters being optimized. Select the Genetic control component (the green dials) to inspect the genetic algorithm settings. Then run the network and observe the improvement of performance over that of the Non-Optimized MLP seen earlier in this demo.
#subtitleTextBox Summary
#mainTextBox This demo has shown the ability of genetic optimization to improve network performance. Genetic optimization can be beneficial any time the network designer is unsure of optimal parameter settings. For more in-depth information regarding genetic algorithms and components, please consult the help files.
