User manual PALISADE NEURALTOOLS 5.5

[. . . ] Guide to Using NeuralTools Neural Network Add-In for Microsoft Excel ® Version 5. 5 February, 2009 Palisade Corporation 798 Cascadilla St. Ithaca, NY USA 14850 (607) 277-8000 (607) 277-8001 (fax) http://www. palisade. com (website) sales@palisade. com (e-mail) Copyright Notice Copyright © 2009, Palisade Corporation. Trademark Acknowledgments Microsoft, Excel and Windows are registered trademarks of Microsoft, Inc. IBM is a registered trademark of International Business Machines, Inc. Palisade, TopRank, BestFit and RISKview are registered trademarks of Palisade Corporation. Welcome to NeuralTools for Excel Welcome NeuralTools gives Microsoft Excel - the industry-standard data analysis and modeling tool - a new, powerful modeling toolset! [. . . ] (If AutoTesting or Auto-Prediction are used, those cases are not included. The reason is that they might have numeric values outside the training range, which could make analysis results more unpredictable. ) 2) For a given category independent variable, for every case the analysis steps through all the valid categories for that variable, and measures the change to the predicted value. (With category prediction there is no numeric predicted value, but there are raw numeric net outputs on which the category prediction is based; those numeric outputs are used by the analysis. ) 3) For a given numeric independent variable, for every case the analysis steps through the range from the minimum to the maximum training value for that variable, measuring the change to the predicted value (or, in the case of category prediction, change to the raw numeric outputs). 46 Command Reference The purpose of the Variable Impact Analysis is not meant to support firm conclusions, like stating with high confidence that a given variable is irrelevant. Instead, it's meant to help in a search for the best set of independent variables: the results of the analysis may be telling us that a given variable looks irrelevant, sufficiently so that it's worth trying to train a net without this variable. The results of a Variable Impact analysis are displayed in the Training Summary report: Reference: NeuralTools Menu Commands 47 Net Configuration Tab The Net Configuration tab in the Training dialog allows you to select the type of neural network that will be trained on your data. You may select a specific net configuration or select a Best Net search where NeuralTools will test a variety of possible configurations to identify the best performing one for you. NeuralTools supports different neural network configurations to give the best possible predictions. For classification/category prediction, two types of networks are available: Probabilistic Neural Networks (PNN) and Multi-Layer Feedforward Networks (MLF). Numeric prediction can be performed using MLF networks, as well as Generalized Regression Neural Networks (GRNN), which are closely related to PNN networks. For more information on the technical aspects of the available network configurations, see the More on Neural Networks section. The Net Configuration tab includes the following: · Type of Net ­ Selects the type of net to be used in training, or alternatively, selects a Best Net search. The Net Configuration tab Options change depending on the type of net selected. In a Best Net Search, NeuralTools tests all checked net configurations, including PNN/GRNN and MLFN nets with node counts in the entered minimum-maximum range. If Store All Trial Nets in New Workbook is selected, you will be able to individually load each tested net (regardless if it was the best performing network) from a workbook and use it for prediction after training is done; a full testing Summary Report for each net will also be available. Command Reference 48 2) PNN/GRNN Net. These net types require no additional options to be selected for training; for this reason this setting is the default when NeuralTools is installed. If your data has numeric output values a GRNN network will be trained and if your data has categorical output values a PNN network will be trained. A Multi-Layer Feedforward Network (MLFN) has one or two hidden layers of nodes. 3) By selecting zero nodes for the second layer it will be eliminated. The most reliable way to find the best configuration of an MLFN net is to use the Best Net Search option instead of the option to train a single MLFN net. If there is not enough time for Best Net Search, it is recommended that the "Number of Nodes" values be left as "Auto-Selected". Reference: NeuralTools Menu Commands 49 Runtime Tab The Runtime tab in the Training dialog allows you to enter stopping conditions for training. If no stopping conditions are selected, training will stop eventually; that period will be relatively short for PNN/GRNN nets, and much longer with MLF nets. One possible approach is to select no stopping conditions and click the Stop button in the training progress dialog when no more time is available for training. With Best Net Search a time limit for training a single net must be defined, to ensure that the search algorithm does not devote too much time to one particular configuration. The three available stopping conditions can be combined, specifying that NeuralTools will stop when any of the conditions are reached. The Stop Training options include: · Train No Longer Than ­ Specifies a fixed time limit for training a single network. [. . . ] So will methods that focus on central tendencies, since the unknown case is closer to the centroid of the square category than to the centroid of the circle category. It will consider the distance of the new case to every training case, giving greater weight to closer cases. The effect of the neighboring square will be outweighed by the circles in the immediate vicinity. More on Neural Networks 89 PNN Architecture A Probabilistic Neural Net is structured as shown in the graph, which assumes there are two independent numeric variables, two dependent categories, and five training cases (three in one category and two in the other): Output Summation Layer (one neuron per category) Pattern Layer (one neuron per training case) Inputs When a case is presented to the net, each neuron in the Pattern Layer computes the distance between the training case represented by the neuron, and the input case. The value passed to Summation Layer neurons is a function of the distance and smoothing factors. [. . . ]