NeuralWorks Predict 3.1

Category Intelligent Software>Neural Network Systems/Tools, Intelligent Software>Fuzzy Logic Systems/Tools and Intelligent Software>Genetic Algorithm Systems/Tools

Abstract NeuralWorks Predict is a complete application development environment for creating and deploying real-time applications for forecasting, modeling, classification and clustering or grouping. This powerful system combines neural network technology with fuzzy logic, statistics and genetic algorithms to find solutions to forecasting, modeling and classification problems automatically. For advanced and expert users, product provides direct access and control of automatic features at a very low level.

Predict requires No prior knowledge of neural networks. With only minimal user involvement it addresses all the issues associated with building robust models from available empirical data. Predict analyzes input data to identify appropriate transforms, partitions the input data into training and test sets, selects relevant input variables, and then constructs, trains, and optimizes a neural network tailored to the problem.

In Microsoft Windows environments Predict can be run either as an add-in for Microsoft Excel to take advantage of Excel's rich data handling and graphing capabilities, or as a command line program that offers powerful batch mode processing. In UNIX and Linux environments, Predict runs as a command line program.

The basic process of building a model with Predict is completely automated. With the Build Wizard activated a short series of dialog boxes provide a step-by-step guide. Five (5) small dialogs determine the layout of data within a spreadsheet. Five (5) more small dialogs determine the complexity of the application and how hard and long Predict should run to produce a solution. Neural network results are written back into the spreadsheet, where a quick graph makes the results suitable for presentation to a wide variety of audiences. The complete system has five (5) main components:

1) The Train/Test Selection component picks out training and test sets for model building. It tries to do this in such a way that the test set is statistically close to the training set. It also allows you to hold back a portion of your data for independent validation of your model.

2) The Data Analysis and Transformation component automatically analyzes data and transforms it into forms suitable for Neural Networks. The types of function that this component performs is to expand categorical data into numeric data, to shape numeric data to get rid of skewness and other undesirable characteristics, to deal with outliers in the data, and to screen out data that contain No information.

3) The Input Variable Selection component uses a genetic algorithm to search for synergistic sets of input variables which are good predictors of the output. Because of the evolutionary nature of the Input Variable algorithm, different initializations of the algorithm will yield different variable sets. You can use this to your advantage to build several models based on different variable sets and combine the outputs of those models. Each model can be thought of as an expert who uses a different set of criteria (the selected variables) to make its decision. There is also an option to do a pre-selection of variables using a Cascaded genetic algorithm approach. This method gives more consistent variable sets by pruning out variables which are consistently rejected by different invocations of the genetic algorithm.

4) The Neural Net component of Predict supports two proprietary non- linear feed-forward constructive algorithms. One of the algorithms is based on a non-linear Kalman filter learning rule and is designed for noisy regression problems. The other is a general-purpose algorithm which is based on an Adaptive Gradient learning rule. Multiple networks may be trained for optimal results. Regularization mechanisms provide good generalization. Eleven (11) different evaluation functions are provided for evaluating test performance during training. Users can trade off speed of learning for comprehensive solutions.

5) The Flash Code component converts the completed model into C, FORTRAN, or Visual Basic (VB) code.

Predict automatically performs all the actions necessary to build a prediction or classification model. A genetic algorithm rapidly builds and evaluates mini-networks to identify Not only which domain inputs are significant, but also the type of transform function that ultimately produces the best network. Then the final neural network is constructed, trained and, and tested. In many situations the resulting network can be deployed immediately.

Predict incorporates six (6) basic transform types and two additional miscellaneous transforms. The basic transform types are:

1) Continuous;

2) Logical;

3) Enumerated Integer;

4) Enumerated String;

5) Fuzzy;

6) Quintile.

The first miscellaneous transform applies to either of the Enumerated transform types. If there are two (2) or more categories that rarely occur, they are combined into a single category labeled ‘Other’. The result of the transform is Tmax if the input field doesn't match any of the categories of the corresponding Enumerated set; otherwise the transform result is Tmin.

The second miscellaneous transform applies to missing or invalid numeric data. If an input field has no data value, or is not a valid numeric value, the result of this transform is Tmax. Otherwise the result of this transform is Tmin.

Continuous transforms that are available in Predict include:

1) Linear (the identity transform);

2) Log (the natural logarithm function);

3) LogLog (the logarithm of the logarithm function);

4) Exp (the exponential function);

5) ExpExp (the exponential of the exponential function);

6) Pwr2 (the square function);

7) Pwr4 (the fourth power function);

8) Rt2 (the square root function);

9) Rt4 (the fourth root function);

10) Inv (the inverse function - 1/x);

11) InvPwr2 (the inverse of the square function);

12) InvPwr4 (the inverse of the fourth power function);

13) InvRt2 (the inverse of the square root function);

14) InvRt4 (the inverse of the fourth root function);

15) Tanh (the hyperbolic tangent function);

16) LnX/(1-X) (the natural logarithm of (x/(1-x)).

Logical transforms that are available in Predict include:

1) Logical;

2) Reverse Logical. These transforms convert continuous data into logical (two-valued) data, by comparing the input value to (Inmin + Inmax)/2.

Enumerated Integer and Enumerated String transforms convert ranges of values into discrete categories. The effect on string (literal) variables is to produce a 1 of N encoding of the variables.

Fuzzy transforms convert input values to "fuzzy" values between 0 and 1, based on a user specified transition point identified by left, center, or right.

The Quintile transform uses piece-wise linear transformations to map input values into 5 bins, such that approximately equal numbers of input records end up in each bin.

System Requirements

On Intel processors, NeuralWorks Predict requires Windows 95 or 98, ME, NT, 2000 or XP; minimum 64 MB memory; 20 MB disk space. Microsoft Excel 97 or later. Contact NeuralWare for details on the Linux and Unix versions supported.

Manufacturer

Manufacturer Web Site NeuralWorks Predict

Price Contact manufacturer.

G6G Abstract Number 30558

G6G Manufacturer Number 101900