Recurrent neural network-based fuzzy logic system

• US 5,606,646 A
• Filed: 06/24/1994
• Issued: 02/25/1997
• Est. Priority Date: 03/24/1993
• Status: Expired due to Term

• Current Assignee: National Semiconductor Corporation (Texas Instruments, Inc.)
• Original Assignee: National Semiconductor Corporation (Texas Instruments, Inc.)
• Inventors: Unal, Faith A., Khan, Emdadur R.
• Primary Examiner(s): Downs, Robert W.
• Assistant Examiner(s): KATBAB, ABDOLLAH
• Application Number:
  US08/265,194
• Time in Patent Office: 977 Days
• Field of Search: 395/24, 395/3, 395/21, 395/22, 395/23, 382/158
• US Class Current: 706/2
• CPC Class Codes: G06N 3/043 based on fuzzy logic, fuzzy membership or fuzzy inference, e.g. adaptive neuro-fuzzy inference systems [ANFIS]
• CPC Class Codes: G06N 7/046 Implementation by means of a neural network neural networks using fuzzy logic G06N3/043
• CPC Class Codes: Y10S 706/90 Fuzzy logic

Abstract

A recurrent, neural network-based fuzzy logic system includes neurons in a rule base layer which each have a recurrent architecture with an output-to-input feedback path including a time delay element and a neural weight. Further included is a neural network-based, fuzzy logic finite state machine wherein the neural network-based, fuzzy logic system has a recurrent architecture with an output-to-input feedback path including at least a time delay element. Still further included is a recurrent, neural network-based fuzzy logic rule generator wherein a neural network receives and fuzzifies input data and provides data corresponding to fuzzy logic membership functions and recurrent fuzzy logic rules.

Claims

1. A neural network-based, fuzzy logic finite state machine comprising:

a neural network which includes a plurality of signal ports for receiving a plurality of input signals, including an external input signal and a present state signal, and providing in accordance therewith a next state signal; and

a feedback path, coupled to multiple ones of said plurality of signal ports, for receiving and time-delaying said next state signal to provide said present state signal.

2. A neural network-based, fuzzy logic finite state machine as recited in claim 1, wherein said neural network comprises:

a first plurality of neurons for receiving and fuzzifying said plurality of input signals to provide a fuzzified plurality of input signals;

a second plurality of neurons, coupled to said first plurality of neurons, for receiving and processing said fuzzified plurality of input signals in accordance with a plurality of fuzzy logic rules to provide a plurality of output signals which corresponds to a plurality of fuzzy logic rule consequents; and

an output neuron layer, coupled to said second plurality of neurons, for receiving and defuzzifying said plurality of output signals and in accordance therewith providing said next state signal.

4. A neural network-based fuzzy logic finite state machine as recited in claim 1, wherein said plurality of signal ports is further for receiving a prior state signal as one of said plurality of input signals.

5. A neural network-based, fuzzy logic finite state machine as recited in claim 4, wherein said feedback path is further for receiving and delaying said next state signal by a plurality of unit time delays to provide said prior state signal.

6. A recurrent artificial neural network for generating pluralities of signals representing a plurality of fuzzy logic rules and a plurality of fuzzy logic membership functions, comprising:

first neural means for receiving a plurality of signals representing input data for an intelligent system and for providing fuzzified data which corresponds to said input data;

second neural means coupled to said first neural means for receiving said fuzzified data and in accordance therewith generating a plurality of membership signals which correspond to a plurality of fuzzy logic membership functions; and

third neural means, including output-to-input feedback means and coupled to said second neural means, for receiving said plurality of membership signals and in accordance therewith recurrently generating a plurality of logic rule signals which represent a plurality of recurrent fuzzy logic rules;

wherein said first, second and third neural means cooperate together by performing a learning process with back-propagation of an output error signal, and wherein said output error signal is propagated serially back from an output of said third neural means through and successively processed by said third, second and first neural means.

3. A neural network-based, fuzzy logic finite state machine as recited in claim 7, wherein said feedback path is for delaying said next state signal by a unit time delay to provide said present state signal.

7. A recurrent artificial neural network as recited in claim 6, wherein said third neural means multiplies selected ones of said received plurality of membership signals.

8. A recurrent artificial neural network as recited in claim 6, further comprising comparison means coupled to said third neural means for receiving therefrom an output signal representing a plant control signal, for receiving an input signal representing a desired control signal, and for comparing said input desired control signal with said output plant control signal and in accordance therewith generating said output error signal.

9. A recurrent artificial neural network as recited in claim 6, further comprising fourth neural means coupled to said third neural means for receiving and summing a plurality of signals therefrom to provide an output sum signal representing defuzzified data.

10. A recurrent artificial neural network as recited in claim 9, further comprising comparison means coupled to said fourth neural means for receiving therefrom said output sum signal, for receiving an input signal representing a desired control signal, and for comparing said input desired control signal with said output sum signal and in accordance therewith generating said output error signal, wherein said output sum signal represents a plant control signal.

11. A recurrent artificial neural network for generating pluralities of signals representing a plurality of fuzzy logic rules and a plurality of fuzzy logic membership functions, comprising:

a first plurality of artificial neurons which receive a plurality of signals representing input data for an intelligent system and provide fuzzified data which correspond to said input data;

a second plurality of artificial neurons, coupled to said first plurality of artificial neurons, which receive said fuzzified data and in accordance therewith generate a plurality of membership signals which correspond to a plurality of fuzzy logic membership functions; and

a third plurality of artificial neurons, including output-to-input feedback for each one thereof and coupled to said second plurality of artificial neurons, which receive said plurality of membership signals and in accordance therewith recurrently generate a plurality of logic rule signals which represent a plurality of recurrent fuzzy logic rules, wherein said third plurality of artificial neurons includes a plurality of output weights corresponding to said plurality of logic rule signals;

wherein said first, second and third pluralities of artificial neurons cooperate together by performing a learning process with back-propagation of an output error signal, and wherein said output error signal is propagated serially back from an output of said third plurality of artificial neurons through and successively processed by said third, second and first pluralities of artificial neurons.

12. A recurrent artificial neural network as recited in claim 11, wherein said third plurality of artificial neurons multiply selected ones of said received plurality of membership signals.

13. A recurrent artificial neural network as recited in claim 11, further comprising a comparator, coupled to said third plurality of artificial neurons, which receives therefrom an output signal representing a plant control signal, receives an input signal representing a desired control signal, and compares said input desired control signal with said output plant control signal and in accordance therewith generates said output error signal.

14. A recurrent artificial neural network as recited in claim 11, further comprising an artificial neuron, coupled to said third plurality of artificial neurons, which receives and sums a plurality of signals therefrom to provide an output sum signal representing defuzzified data.

15. A recurrent artificial neural network as recited in claim 14, further comprising a comparator, coupled to said artificial neuron, which receives therefrom said output sum signal, receives an input signal representing a desired control signal, and compares said input desired control signal with said output sum signal and in accordance therewith generates said output error signal, wherein said output sum signal represents a plant control signal.

16. A computer-implemented method for recurrently generating pluralities of signals representing a plurality of fuzzy logic rules and a plurality of fuzzy logic membership functions, comprising the computer-implemented steps of:

receiving a plurality of signals representing input data for an intelligent system;

generating a plurality of fuzzified data signals representing fuzzified data which corresponds to said input data;

generating in accordance with said plurality of fuzzified data signals a plurality of membership signals which correspond to a plurality of fuzzy logic membership functions;

generating recurrently in accordance with said plurality of membership signals and a plurality of local feedback signals a plurality of logic rule signals which represent a plurality of recurrent fuzzy logic rules; and

performing a learning process with back-propagation of an output error signal in cooperation with said steps of generating said pluralities of fuzzified data signals, membership signals and logic rule signals, wherein said back-propagation of said output error signal is performed following a completion of one iteration of said step of generating said plurality of logic rule signals and is performed successively and prior to subsequent iterations of said steps of generating said pluralities of fuzzified data signals, membership signals and logic rule signals.

17. A method as recited in claim 16, wherein said step of generating recurrently in accordance with said plurality of membership signals and a plurality of local feedback signals a plurality of logic rule signals which represent a plurality of recurrent fuzzy logic rules comprises outputting a plurality of weight signals corresponding to a plurality of output weights associated with a third plurality of artificial neurons, wherein said plurality of weight signals represent said plurality of logic rule signals.

18. A method as recited in claim 17, wherein said step of generating recurrently in accordance with said plurality of membership signals and a plurality of local feedback signals a plurality of logic rule signals which represent a plurality of recurrent fuzzy logic rules further comprises multiplying selected ones of said received plurality of membership signals.

19. A method as recited in claim 16, wherein said step of generating recurrently in accordance with said plurality of membership signals and a plurality of local feedback signals a plurality of logic rule signals which represent a plurality of recurrent fuzzy logic rules comprises multiplying selected ones of said received plurality of membership signals.

20. A method as recited in claim 16, further comprising the steps of:

providing an output signal representing a plant control signal;

receiving an input signal representing a desired control signal; and

comparing said input desired control signal with said output plant control signal and in accordance therewith generating said output error signal.

21. A method as recited in claim 16, further comprising the steps of:

generating a plurality of output signals associated with said plurality of logic rule signals; and

summing said plurality of output signals associated with said plurality of logic rule signals to provide an output sum signal representing defuzzified data.

22. A method as recited in claim 21, further comprising the steps of:

receiving said output sum signal, wherein said output sum signal represents a plant control signal;

receiving an input signal representing a desired control signal; and

comparing said input desired control signal with said output sum signal and in accordance therewith generating said output error signal.

23. A recurrent artificial neural network for defuzzifying a plurality of fuzzy data generated in accordance with a plurality of fuzzy logic rules and membership functions, comprising:

rule-based recurrent neural means for receiving and recurrently processing a plurality of fuzzy signals representing fuzzified data which has been fuzzified in accordance with a plurality of fuzzy logic membership functions to provide in accordance therewith a plurality of recurrent processed fuzzy signals;

output weights means coupled to said rule-based recurrent neural means for receiving and weighting said plurality of recurrent processed fuzzy signals; and

single output neuron means coupled to said output weights means for receiving and summing said plurality of weighted, recurrently processed fuzzy signals to provide an output signal representing defuzzified data.

24. A recurrent artificial neural network as recited in claim 23, wherein said rule-based recurrent neural means multiplies selected ones of said plurality of fuzzy signals.

25. A recurrent artificial neural network as recited in claim 23, wherein said plurality of fuzzy signals corresponds to a plurality of fuzzy logic membership functions.

26. A recurrent artificial neural network as recited in claim 23, wherein said output weights means corresponds to a plurality of fuzzy logic rules.

27. A recurrent artificial neural network as recited in claim 23, further comprising fuzzifier neural means coupled to said rule-based recurrent neural means for receiving and fuzzifying a plurality of data signals to provide said plurality of fuzzy signals.

28. A recurrent artificial neural network for defuzzifying a plurality of fuzzy data generated in accordance with a plurality of fuzzy logic rules and membership functions, comprising:

a plurality of recurrent artificial neurons which receive and recurrently process a plurality of fuzzy signals representing fuzzified data which has been fuzzified in accordance with a plurality of fuzzy logic membership functions to provide in accordance therewith a plurality of recurrently processed fuzzy signals;

a plurality of interneuron weights, coupled to said plurality of recurrent artificial neurons, which receive and weight said plurality of recurrently processed fuzzy signals; and

a single artificial neuron, coupled to said plurality of interneuron weights, which receives and sums said plurality of weighted, recurrently processed fuzzy signals to provide an output signal representing defuzzified data.

29. A recurrent artificial neural network as recited in claim 28, wherein said plurality of recurrent artificial neurons multiply selected ones of said plurality of fuzzy signals.

30. A recurrent artificial neural network as recited in claim 28, wherein said plurality of fuzzy signals corresponds to a plurality of fuzzy logic membership functions.

31. A recurrent artificial neural network as recited in claim 28, wherein said plurality of interneuron weights corresponds to a plurality of fuzzy logic rules.

32. A recurrent artificial neural network as recited in claim 28, further comprising a plurality of fuzzifier artificial neurons, coupled to said plurality of recurrent artificial neurons, which receive and fuzzify a plurality of data signals to provide said plurality of fuzzy signals.

33. A computer-implemented method for defuzzifying a plurality of fuzzy data generated recurrently in accordance with a plurality of fuzzy logic rules and membership functions, comprising the computer-implemented steps of:

receiving a plurality of fuzzy signals representing fuzzified data which has been fuzzified in accordance with a plurality of fuzzy logic membership functions;

recurrently processing said plurality of fuzzy signals to provide a plurality of recurrently processed fuzzy signals;

weighting said plurality of recurrently processed fuzzy signals; and

summing said plurality of weighted, recurrently processed fuzzy signals, in accordance with a single summing function, to provide an output signal representing defuzzified data.

34. A method as recited in claim 33, wherein said step of processing said plurality of fuzzy signals to provide a plurality of processed fuzzy signals comprises multiplying selected ones of said plurality of fuzzy signals.

35. A method as recited in claim 33, wherein said plurality of fuzzy signals corresponds to a plurality of fuzzy logic membership functions.

36. A method as recited in claim 33, wherein said step of weighting said plurality of recurrently processed fuzzy signals comprises processing said plurality of recurrently processed fuzzy signals in accordance with a plurality of fuzzy logic rules.

37. A method as recited in claim 33, further comprising the step of receiving and fuzzifying a plurality of data signals to provide said plurality of fuzzy signals.

38. A recurrent artificial neural network for processing a plurality of input signals as a plurality of fuzzy logic rule antecedents, comprising:

input neural means for receiving and fuzzifying a plurality of input data signals to provide a plurality of fuzzified data signals corresponding to a plurality of fuzzy logic rule antecedents; and

recurrent output neural means coupled to said input neural means for receiving and recurrently multiplying together selected ones of said plurality of fuzzified data signals to provide a plurality of recurrent product signals corresponding to a plurality of fuzzy logic rule consequents.

39. A recurrent artificial neural network as recited in claim 38, further comprising output weights means coupled to said recurrent output neural means for receiving and weighting said plurality of recurrent product signals to provide a plurality of weighted recurrent product signals.

40. A recurrent artificial neural network as recited in claim 39, wherein said output weights means corresponds to a plurality of fuzzy logic rules.

41. A recurrent artificial neural network as recited in claim 39, further comprising neural summing means coupled to said output weights means for receiving and summing said plurality of weighted recurrent product signals.

42. A recurrent artificial neural network for processing a plurality of input signals as a plurality of fuzzy logic rule antecedents, comprising:

a first plurality of artificial neurons which receive and fuzzify a plurality of input data signals to provide a plurality of fuzzified data signals corresponding to a plurality of fuzzy logic rule antecedents; and

a second plurality of artificial neurons, including output-to-input feedback for each one thereof and coupled to said first plurality of artificial neurons, which receive and recurrently multiply together selected ones of said plurality of fuzzified data signals to provide a plurality of recurrent product signals corresponding to a plurality of fuzzy logic rule consequents.

43. A recurrent artificial neural network as recited in claim 42, further comprising a plurality of interneuron weights, coupled to said second plurality of artificial neurons, which receive and weight said plurality of recurrent product signals to provide a plurality of weighted recurrent product signals.

44. A recurrent artificial neural network as recited in claim 43, wherein said plurality of interneuron weights corresponds to a plurality of fuzzy logic rules.

45. A recurrent artificial neural network as recited in claim 43, further comprising an artificial neuron, coupled to said plurality of interneuron weights, which receives and sums said plurality of weighted recurrent product signals.

46. A computer-implemented method for recurrently processing a plurality of input signals as a plurality of fuzzy logic rule antecedents, comprising the computer-implemented steps of:

receiving a plurality of input data signals;

fuzzifying said plurality of input data signals to provide a plurality of fuzzified data signals corresponding to a plurality of fuzzy logic rule antecedents; and

recurrently multiplying together selected ones of said plurality of fuzzified data signals and a plurality of local feedback signals to provide a plurality of recurrent product signals corresponding to a plurality of fuzzy logic rule consequents.

47. A method as recited in claim 46, further comprising the step of receiving and weighting said plurality of recurrent product signals to provide a plurality of weighted recurrent product signals.

48. A method as recited in claim 47, wherein said step of receiving and weighting said plurality of recurrent product signals to provide a plurality of weighted recurrent product signals comprises receiving and weighting said plurality of recurrent product signals in accordance with a plurality of fuzzy logic rules.

49. A method as recited in claim 47, further comprising the step of summing said plurality of weighted recurrent product signals.

50. A recurrent artificial neural network for processing a plurality of input data in accordance with fuzzy logic, comprising:

input neural means for receiving and fuzzifying a plurality of input signals representing numerical data to provide a plurality of fuzzy signals representing fuzzified data; and

recurrent output neural means coupled to said input neural means for receiving and recurrently processing said plurality of fuzzy signals in accordance with a plurality of numerical fuzzy logic rule consequents in the form of singletons corresponding to unit-dimensional data to provide a plurality of numerical signals.

51. A recurrent artificial neural network as recited in claim 50, wherein said plurality of numerical fuzzy logic rule consequents comprise a plurality of neural weights associated with said recurrent output neural means.

52. A recurrent artificial neural network as recited in claim 50, further comprising neural summing means coupled to said recurrent output neural means for receiving and summing said plurality of numerical signals.

53. A recurrent artificial neural network for processing a plurality of input data in accordance with fuzzy logic, comprising:

a first plurality of artificial neurons which receive and fuzzify a plurality of input signals representing numerical data to provide a plurality of fuzzy signals representing fuzzified data; and

a second plurality of artificial neurons, including output-to-input feedback for each one thereof and coupled to said first plurality of artificial neurons, which receive and recurrently process said plurality of fuzzy signals in accordance with a plurality of numerical fuzzy logic rule consequents in the form of singletons corresponding to unit-dimensional data to provide a plurality of numerical signals.

54. A recurrent artificial neural network as recited in claim 53, wherein said plurality of numerical fuzzy logic rule consequents comprise a plurality of neural weights associated with said second plurality of artificial neurons.

55. A recurrent artificial neural network as recited in claim 53, further comprising an artificial neuron, coupled to said second plurality of artificial neurons, which receives and sums said plurality of numerical signals.

56. A computer-implemented method for recurrently processing a plurality of input data in accordance with fuzzy logic, comprising the computer-implemented steps of:

receiving a plurality of input signals representing numerical data;

fuzzifying said plurality of input signals to provide a plurality of fuzzy signals representing fuzzified data; and

recurrently processing said plurality of fuzzy signals and a plurality of local feedback signals in accordance with a plurality of numerical fuzzy logic rule consequents in the form of singletons corresponding to unit-dimensional data to provide a plurality of numerical signals.

57. A method as recited in claim 56, further comprising the step of summing said plurality of numerical signals.