The bottom line is that a neural network is a classifier of a set of values of a fixed set of independent variables. The classification can be a decision, a diagnosis, or a control action. Stated in a different way, a neural network maps a fixed set of independent variables to a dependent variable: IF the “temperature” is moderate and the “day” is the weekend and the “lawn” has been mowed, THEN “go to the golf club” (where the IF conditions are the independent variables (temperature, day, lawn) and the THEN classification (go to the golf club) is the dependent variable. Another set of conditions, another classification (say, “clean the house”). The values of the independent variables can be analog (23, 39, 8), digital (1/0), or categorical (“color”, or “month of the year”, or etc.).

The table below provides the basis for explaining all the examples of how the neural network is used in this report. The number of columns can be up to a thousand, while the number of rows can be in the tens of thousands. The larger numbers of course imply longer times to train the network.

The letters above (A,B,C,D,E,F,G) each represent a parameter (independent variable) such as the temperature, the air pressure, the the water level, the velocity, the $ amount of the accounts receivable, the number of rooms on a house, the $ amount of the cost of sales, etc., etc. They could be anything to which one could assign a value, such as 72 degrees, 12 psi, 45 mph, etc. A parameter such as “color” would have separate columns: A=red, B=green, C=blue, etc. The column A could in turn have a value of 1-100 to represent various shades. The numbers to the right of the table (1,2,3,4,5) represent the rows of the table. As explained later, each row is an “exemplar” of a given set of conditions. An example of an exemplar would be a syndrome of a disorder (syndrome equals a group of symptoms that together are characteristic of a specific condition). These exemplars are generated by a group of experts of a given field of study (bank loan specialists, for example). This process by the experts is the creative part. Thus, this group would arrive at a consensus that for a given set of values for a given set of parameters, the loan applicant is “good”. For another set of values, the applicant is “excellent”, or “poor” , or whatever. The parameters are independent variables. So the experts create this table of exemplars, with each set of values representing a unique outcome: “excellent”, “good”, “fair”, or “poor”. Each exemplar is a unique IF-THEN statement: IF A=27 and B=5760 and C=170 and etc. THEN the decision/diagnosis/control action is #42. The more exemplars, the more accurate will be the “THEN” part of the statement. Also, there could be multiple “#42’s” representing multiple sets of “IF” values. Think of the exemplar as being a vector pointing in some direction from the center of a sphere. Then a vector in the x=5, y=16, z=20 direction could represent the same decision/diagnosis/control action as a vector in the x=17, y=2.z-14 direction.

So, if a given unique set of values for a given unique set of parameters (variables A, B, C, …..) pertains, then decision/diagnosis/control action “#42”, say. If another, different, unique set of values (for the same set of variables A, B, C,…..), then decision/diagnosis/control action “#9”, say. (The above values can be obtained from a set of sensors, either manually recorded or automatically sent to the neural network (computer)). The problem to be resolved at the outset of the design of the system is “what set of conditions (a12, b5, c2, etc.) represent a given decision/diagnosis/control action (#1, #2, #3, etc.)?”. This set of exemplars (the table) can be considered a model of the system at hand. For instance, one way of buiding this model is to use a history of an equipment’s malfunctions as the raw data source. That is, whenever an equipment malfunctions, record the values of the pre-determined set of parameters (A, B, C, …..). For an electronic circuit board, these could be “Test Point A, Test Point B, etc.”. Then when a circuit board has a malfunction, the technician takes readings at TP A,B,C, etc., applies these readings to the NN, and the NN interpolates the table to determine the closest match to one of the exemplars. This found exemplar then identifies the diagnosis “#71”, say. Each reading of the sensor values, at sampled times, is a “signature” of that particular malfunction.

Neural Network Applications Methodology

1. A Neural Network Lookup Table methodolgy is a solution looking for a problem to solve
2. The “Problem” must be in the form where a decision, a diagnosis, a control action is a function of a set of conditions as represented by interrelated, multiple-input variables (exemplars—>the independent variables).
3. The “Solution” is to teach a neural network a basic set of these input-output combinations (a1, b1, c1, etc.—>#1; a2, b2. c2, etc.—>#2, where a1,a2,b1,b2,c1,c2, are arbitrary values). Once taught, the NN can accept any arbitrary set of values a,b,c, etc. (the input condition) and “instantaneously”provide the corresponding output.
4. Therefore, the effort takes the form of finding those problems that have a unique output for a unique set of interrelated, multiple inputs.

Some Specific Applications Are

Corrective Maintenance (machine, electronic, medical diagnosis)

Preventive Maintenance (machine, electronic, medical diagnosis)

Supervisory Control

Decision Support Systems (DSS)

(Implicit in the above is the effort to determine the list of independent variables that will, together, detect the source of the malfunction or disorder——or incipient malfunction/disorder).

(The HNeT computer is able to easily handle thousands of Variables each of which can handle hundreds of Values. The Learning of the data set and the Results (providing the Decision) are available in a matter o seconds. The HNeT Computer is ‘in another universe’.)

See Neurodecisions 2.