Greetings from NeuroDimension!
Makers of NeuroSolutions, the Neural Network Simulation Environment.
This issue of the newsletter highlights several new developments at NeuroDimension that will help you to get the most out of NeuroSolutions, as well as tips and hints for designing neural networks from the professionals at ND!
In this issue you’ll find:
What’s New?
Designing Neural Networks
NeuroSolutions Tip Box
Did You Know?
Customer Spotlight
Note: You are receiving this newsletter because you requested to stay informed concerning new developments at NeuroDimension. If you would like to stop receiving these newsletters, please see the bottom of this page for removal instructions.
What’s New?
Download DLLs from the ND Web Site
A new section has been added to our web site to provide users of NeuroSolutions with access to new component DLLs (with source code) as soon as they become available. The engineers at NeuroDimension are continually developing new DLL components to further enhance the package. In addition, many of our Developers level customers have written DLLs for their needs, and are encouraged to submit the code for any DLLs that could be useful to others. Visit the new DLL Download section at: http://www.nd.com/dll.htm
NeuroDimension wants NeuroSolutions to be your standard development platform for new neural network algorithms. Stop by today to check out this new section and download any DLLs that you may find useful. Also, if you have created some interesting non-proprietary DLLs, we encourage you to consider sharing them with the NeuroSolutions community.
Printable Documentation Now Available
Each copy of NeuroSolutions ships with a small (50-page) "Getting Started" manual. Because the remainder of the NeuroSolutions documentation is so extensive (over 700 pages) and rapidly changing, it is not practical for us to offer a printed version. However, there are times when one might want to have a printed chapter to read at times when their computer is not available. For this reason, we have recently converted the contents of the entire online help to a MS Word 97 document and made it available for download from our web site: http://www.nd.com/support/help.htm
Coming Soon - Genetic Server / Genetic Library
NeuroDimension will be releasing two new products in the coming weeks. Each of these products provides a general purpose application programmer’s interface (API) for genetic algorithm (GA) design. The Genetic Server is an Active X component designed to be used within Visual Basic or software supporting Visual Basic for Applications (such as Microsoft Excel, Microsoft Access, etc.). The Genetic Library is a C++ library designed to be used within Visual C++. If you’ve been waiting to try out GA on one of your problems, now is the time. With their intuitive APIs, the Genetic Server and Genetic Library make using genetic algorithms easy. For more information, see http://www.nd.com/products/genetic.htm
Please send a blank email to genetic@nd.com to be notified when these products are officially released.
Designing Neural Networks
This Month: How do you select the size of your network?
A neural network that is too small will not be sufficiently powerful to solve the problem. A network that is too large will not only train slower, but may produce inferior results -- especially in the test set. This is closely related to the generalization problem that is famous in the neural network literature.
A neural network is a parameterized system, where the weights are the adjustable parameters. The more weights, the more powerful the network -- and the more training data you need to adequately train the network. The number of weights in a network is determined by the number of hidden layers, the number of PEs in each layer, the number of inputs, and the number of memory taps. The number of weights in each component in NeuroSolutions is shown on the "Soma" page of the inspector.
The rule of thumb is that you need 10 exemplars of data for each weight in the network. So, if you have only 100 exemplars of data for an MLP, you should probably use only a single hidden layer with only a few PEs. With very large data sets, however, you don't necessarily need a very large network. Always start small and grow the system by adding PEs or hidden layers. Select the smallest model that is not significantly improved with additional PEs, hidden layers, or memory taps. This will provide the system with the best generalization and the fastest training times.
NeuroSolutions Tip Box
This Month: Unsupervised Networks
A common question addressed to our technical support staff is "Why does the NeuralWizard ask for desired data when I selected an unsupervised network?" The answer to the question is that the unsupervised neural models built by the NeuralWizard (SOFM and PCA) are actually hybrid unsupervised/supervised networks. The unsupervised segment works as a preprocessor; i.e., it extracts features from the input data. These features are then fed to the input of a standard MLP.
As demonstrated within the NeuroSolutions demos, the software does support pure unsupervised models. One way to construct your own unsupervised model is to build a hybrid network with the NeuralWizard and delete the supervised portion of the network. To do this, we have created a NeuroSolutions macro that can be downloaded from the ND web site. Here are the basic steps:
Another way to construct a pure unsupervised network is to run one of the unsupervised demos until it builds a network similar to the one that you want. From there, simply replace the demo data with your own. To do this you may need to remove the Function component and replace it with a File component, then specify your data file(s) within the component's inspector.
Did You Know?
This Month: Custom Solution Wizard – Reading Data from Files
It is very easy to read data from a structured file (Comma Delimited, Tab Delimited, Fixed Width, etc.) for input into a Custom Solution Wizard DLL called from Visual Basic or Visual Basic for Applications. The key is to use Microsoft DAO. To use DAO, you must first add a reference to the DAO Object Library from the References menu item within your VB / VBA development environment. The next step is to create a schema.ini file that describes the format of your data and place this file in the same location as your data file. After performing these steps, you will be able to access your data file like a typical database using OpenDatabase, OpenRecordset, MoveNext, GetRows., etc.
For a complete discussion on using DAO to read data from text files, download the updated Custom Solution Wizard help file from http://www.nd.com/support/help.htm. See the "Reading Data from Text Files" topic. To get to this topic, open the updated help file, then from the Index tab, type in "Text files" and click "Display". When the "Topics Found" dialog appears, click "Display" again.
Customer Spotlight
This Month: Predicting Contingency Cost in Construction Management
Akinsola, A.O., School of Engineering and Built Environment, University of Wolverhampton
The construction industry has been consistently criticized for poor performance in attaining its clients' requirements. Time and cost overruns are very common and accepted as an inevitable part of construction. These overruns are a major cause of disruption, delay, disputes and excess cost. Yet no empirical method or tool, quantitative or otherwise, is available for managing or controlling them. The conventional approach is to include a percentage of the project cost as contingency in the pre-contract budget for their occurrences. The allocated contingency based on this method is largely judgmental and arbitrarily allocated, which is often overly simplistic and unrealistic. NeuroSolutions is used to implement a three-layer MLP neural network model to better predict the total contingency cost allowance for variations on a given construction project.
This is just an abstract of the application summary. The entire summary is available at: http://www.nd.com/application%20summaries/appsum-manag.htm
Want to have your solutions spotlighted? We strongly encourage our customers to send their 1-2 page application summaries to submissions@nd.com so that we may post them on our web site at: http://www.nd.com/appliactionsum.htm. In each newsletter, we’ll spotlight a new solution and include a link for people to get more information.
Comments or Suggestions?
We appreciate your feedback! Please send us your comments or suggestions concerning this newsletter, our web site, or part of the NeuroDimension product line. Write to us at: feedback@nd.com
This issue and previous issues of this newsletter are available on the NeuroDimension web site at: http://www.nd.com/mailinglist.htm
If you would prefer not to receive these newsletters or subsequent product updates from NeuroDimension, please reply to this letter with the subject heading changed to the word REMOVE.
Thank you again for your support of NeuroDimension products!