# California Scientific

California Scientific
1000 SW Powell Ct
Oak Grove, MO 64075

# Setting the Min/Max values correctly

NetMaker will set these values for you, but the chosen values will be arbitrarily set and +/- 1.7 standard deviations. This is often not a very good choice. The problem here is that information and data are two very different concepts. Here's how to tell what's right for you:

The thing you're trying to do is emphasize the important parts of your data, while neglecting the irrelevant parts. Suppose you have yearly income's for several hundred average people. In the USA, average personal income is about \$32,000 per year, and the standard deviation is about \$8,000 per year: two thirds of everyone makes between about \$24,000/year and \$40,000/year.

If you were trying to build a neural network to help a bank select good candidates for credit card offers, you would not be very interested in average people. People making \$32,000 per year don't tend to spend very much on credit cards. You would be looking for people making \$40,000/year to \$80,000/year - people who are 1 to 3 standard deviations above average. For this problem, your best min/max values would likely be 40,000 and 80,000.

If you were trying to build a network to predict people who would likely need food stamp assistance in the next 6 months, you would probable be most interested in people making \$5,000/year to \$20,000/year. A person make this yearly income is very vulnerable to losing a job, or even losing a couple of weeks pay due to illness. People making \$32,000 per year and up are unlikely to be so vulnerable to such problems. In this case, you'd probably set your min/max values at 5,000 and 20,000 - 1.5 to 3 standard deviations below average.

Finally, if you were trying to see if parent's income was a valid predictor for a child's school performance, you'd be looking at mostly average kids from average homes, and it's likely the default settings of 20,000 and 44,000 would be reasonable for your problem.

So, with the same basic data - personal yearly income - three different problems leads us to three very different appropriate choices for min/max values. In the case of your data, you must ask yourself if the interesting information is near the average, far above the average, or far below the average. If you're not sure, you could try giving the same data to BrainMaker in three columns with different min/max sets. After training a network, you can then check to see which column was most important and thereby learn where the interesting information was.