Any database used with VisiRex requires one field to be set a primary key. We normally set a counter field (e.i. row number) as the primary key. A bug has been confirmed that VisiRex will not operate correctly if the primary key field is a text field. This problem will be fixed in the next release. For now, please ensure your primrary key is a numeric field.

Hello Andres,

We may include a C module in future versions of the Software Developer's Kit. The SDK currently produces Basic nested IF statements which should convert easily to the C language. A description of the VisiRex SDK is available here.

- Clay

_________________________
Clay Essex
Senior Programmer
CorMac Technologies Inc.
essexc@cormactech.com

Suggestions for structuring data from horse races are discussed on the NeuNet horse racing page.
One suggestion is to show the "fallback" from best in race for each factor. Another suggestion is to use "pairwise matching" where many horses in a single race are considered as multiple two-horse races.

Hello Neil,
Thanks for the excellent suggestion. It should be easy for us to export the text tree in its current state of expansion as text to the clipboard. The text could then be pasted into any word processor or text editor. We will plan this feature for our next upgrade! For now you can print the tree to a seperate page and refer to it from your document. Or you could use a sceen-grab program to take a graphic snapshot of the tree and place into your Word document. Or you could use Viseo to place the flowchart into a Word document.

Hello Robin,
I have found several references to CART in Quinlan's Book "C4.5 Programs for Machine Learning" (ISBN 1-55860-238-0). VisiRex is based on a highly optimized version of Quinlan's C4.5 algorithm.
Page 15: "...statistics has given rise to a multitude of classification methods... How do these methods differ form those explored in machine learning? In some cases there are no differences: CART.., a well-respected system for building decision trees was developed by statisticians. As a general rule, however, statistical techniques tend to focus on tasks in which all the attributes have continuous or ordinal values."
Page 30-33: "If... a relevent attribute is unknown... the system explores all possible outcomes and combines the classifications arithmetically... a classification is a class distribution rather than a single class... the class with the highest probability is assigned the predicted class... It is interesting to compare this approach... with CART... CART finds another test called a surrogate split whose partition of the training cases is most similar to the original test... It depends on finding suitable surrogate splits, which seems unlikely in domains where tests having different numbers of outcomes and where there is little correlation between attributes."
Page 63-64: "The original CLS selects one distinguising value of the splitting attribute for one branch and lumps all the other values together for the other branch. CART also uses binary splits but selects a group of values for both branches. Rather than forcing binary splits, C4.5 associates a value group with each of a variable number of branches."