Browse Prior Art Database

# Table-Driven Parameter Normalization for Expert Systems

IP.com Disclosure Number: IPCOM000102443D
Original Publication Date: 1990-Nov-01
Included in the Prior Art Database: 2005-Mar-17
Document File: 3 page(s) / 123K

IBM

## Related People

Truelson, RW: AUTHOR

## Abstract

Disclosed is a method for deriving the normalized value of an input parameter. This method constructs a customized monotonically-increasing, continuously differentiable normalization function from a table entry containing two threshold values. The method has general application to expert systems and artificial intelligence.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 51% of the total text.

Table-Driven Parameter Normalization for Expert Systems

Disclosed is a method for deriving the normalized value
of an input parameter.  This method constructs a customized
monotonically-increasing, continuously differentiable normalization
function from a table entry containing two threshold values.  The
method has general application to expert systems and artificial
intelligence.

In expert systems, there is frequently a need to characterize
some input parameter.  Sometimes it is sufficient to characterize it
as either "high" or "low", but in other cases a much finer
characterization is required. This can be done with a mathematical
function that converts the raw value of the parameter to a value in
some standard range ("normalized value"); unfortunately, such
functions are difficult to code and maintain.  Another approach is to
maintain a table of thresholds and values they translate to (in
effect, a mathematical step function); however, there is a trade-off
between table complexity and information loss due to granularity.
This invention combines the ease of coding and maintenance of a
simple table with the advantages of specially constructed continuous
normalization functions.

The normalization procedure consists of the following steps:
1) read the parameter's upper and lower thresholds and, optionally,
the range of possible input values from a table,  2) construct a
normalization function from the upper and lower thresholds and range
of input values,  3) read in the parameter value to be normalized,
and 4) calculate the normalized value by plugging the raw value into
the normalization function.  The key feature of this invention is the
second step, i.e., the construction by the system of a useful
normalization function for any arbitrary set of thresholds and range
of input values.

Without loss of generality, the normalization procedure is
designed to yield a value in the range of 0 to 1; a programmer who
desires to normalize to a different range of values can easily do so
by multiplying a value in the 0 to 1 range by an appropriate constant
and offsetting it as necessary.  A useful normalization function
should have several characteristics.  It should be continuous, so
that raw values that are close to each other will convert to
normalized values that are close.  It should be monotonically
increasing, so that each raw value converts to a unique normalized
value and for each normalized value, there is some unique raw value
that converts to it.  It is desirable that the lowest possible input
value normalize to 0, and the highest possible input value normalize
to 1.  It is also desirable that the high and low thresholds
normalize to some constant values between 0 and 1, i.e., if different
parameters in different ranges are being normalized, the respective
high and low thresholds will always normalize to the same value.
Finally, it should have highest gain in the region of greatest
inte...