Method for structuring an expert system utilizing one or more neural networks

• US 5,455,890 A
• Filed: 09/30/1993
• Issued: 10/03/1995
• Est. Priority Date: 09/30/1993
• Status: Expired due to Term

• Current Assignee: Motorola Solutions, Inc.
• Original Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Inventors: Wang, Shay-Ping
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): HAFIZ, TARIQ R
• Application Number:
  US08/129,275
• Time in Patent Office: 733 Days
• Field of Search: 395/11, 395/3, 395/21, 395/22, 395/900, 395/50, 395/61
• US Class Current: 706/16
• CPC Class Codes: G06N 3/042 Knowledge-based neural networks; Logical representations of neural networks
• CPC Class Codes: Y10S 128/925 Neural network

Abstract

Neural networks learn expert system rules, for either business or real-time applications, to improve the robustness and speed of execution of the expert system. One or more neural networks are constructed which incorporate the production rules of one or more expert systems. Each neural network is constructed of neurons or neuron circuits each having only one significant processing element in the form of a multiplier. Each neural network utilizes a training algorithm which does not require repetitive training and which yields a global minimum to each given set of input vectors.

Claims

1. A method for implementing production rules of an expert system using a neural network, said method comprising the steps of:

(a) defining a plurality of inputs and system outputs for said expert system;

(b) defining at least one group of production rules which define a relationship between one of said system outputs and at least one of said plurality of inputs;

(c) defining said neural network, said neural network having at least one network output, said neural network being responsive to said at least one of said plurality of inputs corresponding to said at least one group of production rules, and said neural network comprising a plurality of neurons;

(d) computing the weights of said neural network using training examples derived from the at least one group of production rules;

(e) expressing said neural network by a polynomial expansion; and

(f) computing said at least one network output by substituting said at least one of said plurality of inputs into said polynomial expansion;

wherein said one of said system outputs is based on said at least one network output.

2. The method recited in claim 1, wherein step (f) includes the substep of:

generating said at least one network output as an interpolated value.

3. The method of claim 2, wherein at least one of said system outputs is quantized into any one of a number of N quantified values, N being a positive integer, and said quantified values having a minimum value and a maximum value.

4. The method of claim 1, wherein step (a) further comprises the substep of normalizing said inputs.

5. The method of claim 4, wherein said normalizing substep is performed by transforming the inputs into values between -1 and +1.

6. The method of claim 4, wherein said normalizing substep is performed by transforming the inputs into absolute values between 0 and 1.

7. The method of claim 1, wherein in step (a) each input is quantized into any one of a number of N quantified values, N being a positive integer.

8. The method of claim 1, wherein step (a) further comprises the substep of normalizing at least one of the system outputs.

9. The method of claim 8, wherein said normalizing substep is performed by transforming said at least one system output into a value between -1 and 1.

10. The method of claim 8, wherein said normalizing substep is performed by transforming said at least one system output into an absolute value between 0 and 1.

11. A method for implementing an expert system with a plurality of neural networks, said method comprising the steps of:

(a) defining a plurality of subsystems for a problem, each subsystem having at least one input and at least one subsystem output;

(b) defining a group of production rules for each of said subsystems which define a relationship between said at least one subsystem output and said at least one input;

(c) linking said subsystems together to form said expert system;

(d) defining said plurality of neural networks, each of said neural networks corresponding to a respective subsystem, each of said neural networks responsive to said at least one input of said respective subsystem and having a at least one network output, each of said neural networks comprising a plurality of neurons;

(e) computing the weights of said plurality of neural networks using training examples for each of said neural networks derived from the group of production rules for said respective subsystem; and

(f) expressing each of said neural networks by a polynomial expansion; and

(g) computing said at least one network output from at least one of said neural networks by substituting said at least one input into said polynomial expansion corresponding to said at least one of said neural networks;

wherein said at least one subsystem output is based on said at least one network output.

12. The method recited in claim 11, wherein step (g) includes the substep of:

generating said at least one network output as an interpolated value.

13. The method of claim 12, wherein said at least one subsystem output is quantized into any one of a number of N quantified values, N being a positive integer, and said quantified values having a minimum value and a maximum value.

14. The method of claim 11, wherein step (a) further comprises the substep of normalizing said at least one input.

15. The method of claim 14, wherein said normalizing substep is performed by transforming said at least one input into a value between -1 and +1.

16. The method of claim 14, wherein said normalizing substep is performed by transforming said at least one input into an absolute value between 0 and 1.

17. The method of claim 11, wherein in step (a) said at least one input is quantized into any one of a number of N quantified values, N being a positive integer.

18. The method of claim 11, wherein step (a) further comprises the substep of normalizing said at least one subsystem output.

19. The method of claim 18, wherein said normalizing substep is performed by transforming said at least one subsystem output into a value between -1 and +1.

20. The method of claim 18, wherein said normalizing substep is performed by transforming said at least one subsystem output into an absolute value between 0 and 1.

21. A method for implementing an expert system with a plurality of neural networks comprising the steps of:

(a) defining a plurality of subsystems for a problem, each subsystem having at least one input and at least one subsystem output;

(b) defining a table for each of said subsystems, said table comprising at least one entry defining a relationship between said at least one subsystem output and said at least one input;

(c) linking said subsystems together to form said expert system;

(d) defining said plurality of neural networks, each of said neural networks corresponding to a respective subsystem, each of said neural networks responsive to said at least one input of said respective subsystem and having at least one network output, each of said neural networks comprising a plurality of neurons;

(e) computing the weights of said neural networks using training examples for each of said neural networks derived from the table for said respective subsystem;

(f) expressing each of said neural networks by a polynomial expansion; and

(g) computing said at least one network output by substituting said at least one input into said polynomial expansion corresponding to one of said neural networks;

wherein said at least one subsystem output is based on said at least one network output.

22. The method recited in claim 21, wherein step (g) includes the substep of:

generating said at least one network output as an interpolated value.

23. The method of claim 22, wherein said at least one subsystem output is quantized into any one of a number of N quantified values, N being a positive integer, and said quantified values having a minimum value and a maximum value.

24. The method of claim 21, wherein step (a) further comprises the substep of normalizing said at least one input.

25. The method of claim 24, wherein said normalizing substep is performed by transforming said at least one input into values between -1 and +1.

26. The method of claim 24, wherein said normalizing substep is performed by transforming said at least one input into absolute values between 0 and 1.

27. The method of claim 21, wherein in step (a) said at least one input is quantized into any one of a number of N Quantified values, N being a positive integer.

28. The method of claim 21, wherein step (a) further comprises the substep of normalizing said at least one subsystem output.

29. The method of claim 28, wherein said normalizing substep is performed by transforming said at least one subsystem output into a value between -1 and +1.

30. The method of claim 28, wherein said normalizing substep is performed by transforming said at least one subsystem output into an absolute value between 0 and 1.

31. A method for utilizing an expert system implementable with a plurality of neural networks, said method comprising the steps of:

(a) defining a plurality of subsystems for a problem, each subsystem having at least one input and at least one output;

(b) defining a group of production rules for each of said subsystems which define a relationship between said at least one output and said at least one input;

(c) linking said subsystems together to form said expert system;

(d) defining said plurality of neural networks, one for each corresponding group of production rules, each of said networks comprising a plurality of neurons;

(e) computing the weights of said neural networks using training examples for each of said neural networks derived from the corresponding group of production rules;

(f) expressing each of said neural networks by a polynomial expansion;

(g) distributing a set of input values derived from said at least one input to at least one of said neural networks; and

(h) said at least one neural network calculating an output value for said set of input values by substituting said set of input values into said polynomial expansion corresponding to said at least one of said neural networks.

32. The method recited in claim 31, wherein step (h) includes the substep of:

generating said output value as an interpolated value.

33. The method of claim 31, wherein step (f) further comprises the substep of normalizing said input values.

34. The method of claim 33, wherein said normalizing substep is performed by transforming said inputs into values between -1 and +1.

35. The method of claim 33, wherein said normalizing substep is performed by transforming said inputs into absolute values between 0 and 1.

36. The method of claim 31, wherein in step (a) said at least one input is quantized into any one of a number of N quantified values, N being a positive integer.

37. The method of claim 31, wherein step (a) further comprises the substep of normalizing said at least one output.

38. The method of claim 37, wherein said normalizing substep is performed by transforming said at least one output into a value between -1 and +1.

39. The method of claim 37, wherein said normalizing substep is performed by transforming said at least one output into an absolute value between 0 and 1.

40. The method of claim 31, wherein said output value is quantized into any one of a number of N quantified values, N being a positive integer, and said quantified values having a minimum value and a maximum value.