Neural network having expert system functionality

• US 5,398,300 A
• Filed: 12/30/1991
• Issued: 03/14/1995
• Est. Priority Date: 07/27/1990
• Status: Expired due to Term

• Current Assignee: Fair Isaac Corporation
• Original Assignee: HNC Incorporated
• Inventors: Levey, Curt A.
• Primary Examiner(s): Downs, Robert W.
• Application Number:
  US07/814,179
• Time in Patent Office: 1,170 Days
• Field of Search: 395/20, 395/21, 395/22, 395/51, 395/61, 395/75, 395/76, 395/11, 395/23
• US Class Current: 706/16
• CPC Class Codes: G06N 3/042 Knowledge-based neural networks; Logical representations of neural networks
• CPC Class Codes: G06N 5/022 Knowledge engineering; Knowledge acquisition
• CPC Class Codes: G06N 5/045 Explanation of inference; Explainable artificial intelligence [XAI]; Interpretable artificial intelligence

Abstract

A method for performing a variety of expert system functions on any continuous-state feedforward neural network. These functions include decision-making, explanation, computation of confidence measures, and intelligent direction of information acquisition. Additionally, the method converts the knowledge implicit in such a network into a set of explicit if-then rules.

Claims

1. A computer-implemented process for determining high and low bounds on states of output processing elements (PEs) and any hidden PEs in a continuous-state feedforward neural network additionally having input PEs and that is stored in a computer, wherein said bounds serve as limits on the possible range of each such state, said computer comprising a CPU and an input device coupled to said CPU, said CPU comprising an inference module, said neural network being fed into said CPU via said input device, said process comprising commanding the inference module to execute the steps of:

determining upper and lower bounds on states of the input PEs using input variable distribution information that has been fed into said CPU via said input device;

in succession, computing the bounds on the states of any hidden PEs from the bounds on the input PEs and the computed bounds on the states of any hidden PEs from hidden layers of PEs closer to the input PEs; and

computing the bounds on the states of the output PEs from the bounds on the states of the input PEs and the bounds on the states of any hidden PEs.

3. A computer-implemented process for analyzing the operation of a continuous-state feedforward neural network that is stored in a computer, said computer comprising a CPU, said CPU comprising an inference module, an explanation module, and an intelligent knowledge acquisition module, said neural network being stored within the CPU, where the network includes a layer of input processing elements (PEs), a layer of output PEs, and any number of layers of hidden PEs, where the input PEs are representative of input variables and the output PEs are representative of output variables, said process comprising commanding the inference module to perform the steps of:

determining a tentative decision for each output variable based upon states of the output PEs in the neural network;

determining a final conclusion for each output variable using the process of claim 1 where, as values of input variables become known, the tentative decision eventually becomes irreversible; and

computing confidence measures for each tentative decision and final conclusion;

commanding the explanation module to perform the steps of;

explaining each tentative decision by computing the contribution made towards the decision by each input variable whose value is known;

explaining each final conclusion by computing the contribution made towards reaching the conclusion by each input variable whose value is known; and

converting knowledge implicit in the neural network into an explicit set of if-then rules; and

commanding the intelligent knowledge acquisition module to perform the step of determining potential influence on a tentative decision by each input variable whose value is unknown.

5. Apparatus for determining when a final conclusion can be made regarding the value of an output variable in a continuous-state feedforward neural network, despite some number of unknown input states, said apparatus comprising:

a computer comprising a CPU, an input device coupled to the CPU, and an output device coupled to the CPU;

wherein the neural network is fed into the CPU via the input device, and the CPU comprises an inference module comprising;

means for making a tentative decision for the output variable by executing the neural network;

coupled to said making means, means for determining high and low bounds on states of the output variable using the process of claim 5, wherein the process of claim 5 is executed on said CPU;

coupled to the determining means, means for using the high and low bounds on the output states to determine if the tentative decision for the output variable can ever change; and

coupled to the using means, means for declaring a final conclusion if the tentative decision cannot change and for sending a signal announcing the final conclusion to the output device.

2. A computer-implemented process for determining when a final conclusion can be made regarding the value of an output variable in a continuous-state feedforward neural network that is stored in a computer, despite some number of unknown input states, wherein said computer comprises a CPU and an output device coupled to said CPU, said CPU comprises an inference module, and said neural network is stored within said CPU;

said process comprising commanding the inference module to execute the steps of;

making a tentative decision for the output variable by executing the neural network;

determining high and low bounds on states of the output variable using the process of claim 5;

using the high and low bounds on the output states to determine if the tentative decision for the output variable can ever change; and

sending a signal to the output device declaring a final conclusion if the tentative decision cannot change.

4. A computer-implemented process for determining when a final conclusion can be made regarding the value of an output variable in a continuous-state feedforward neural network that is stored in a coprocessor coupled to a digital computer, despite some number of unknown input states, wherein said coprocessor comprises a CPU and an output device coupled to said CPU, said CPU comprises an inference module, and said neural network is stored within said CPU;

said process comprising commanding the inference module to execute the steps of;

making a tentative decision for the output variable by executing the neural network;

determining high and low bounds on states of the output variable using the process of claim 5;

using the high and low bounds on the output states to determine if the tentative decision for the output variable can ever change; and

sending a signal to the output device declaring a final conclusion if the tentative decision cannot change.

6. A computer-implemented process for computing a measure of certainty for a decision corresponding to the value of an output variable in a continuous-state feedforward neural network that is stored in a computer, said computer comprising a CPU and an output device coupled to the CPU, said CPU comprising an inference module, said neural network being stored with the CPU, said process comprising directing the inference module to perform the steps of:

determining network output states given a set of input variables some of whose values are known and others of whose values are unknown;

determining upper and lower bounds on the output states using the process of claim 5;

comparing the bounds with a preselected conclusion criterion;

determining a certainty of the decision by quantitatively measuring a closeness of the bounds to the conclusion criterion; and

sending a signal to the output device corresponding to said certainty.

7. A computer-implemented process for determining a minimal subset of current input values that is sufficient to support a conclusion condition in a continuous-state feedforward neural network that is stored in a computer, said computer comprising a CPU and an output device coupled to the CPU, said CPU comprising an inference module and an explanation module, said neural network being stored in said CPU, said process comprising directing the inference module to perform the steps of:

determining output states of the neural network based upon a set of input variables, some of whose values are known and others of whose values are unknown;

determining upper and lower bounds on the output states using the process of claim 5; and

determining any final conclusion by comparing the bounds with a preselected conclusion criterion; and

directing the explanation module to perform the steps of;

if a conclusion is possible, iteratively toggling each input variable whose value is known into an unknown condition, thereby measuring a relative importance of the toggled input variable in terms of a change in the output bounds;

defining said minimal subset to be those input variables that are just sufficient to reach the conclusion condition, filling said minimal subset with input variables in descending order of relative importance; and

displaying said minimal subset on said output device.

10. A computer-implemented process for explaining the reasons for a decision made by a continuous-state feedforward neural network that is stored in a computer by measuring the contribution to that decision made by each input variable having a known value, said computer comprising a CPU, an input device coupled to the CPU, and an output device coupled to the CPU;

said neural network having been fed into the CPU via the input device;

said CPU comprising an inference module and a decision explanation module;

said process comprising directing the inference module to execute the steps of;

determining network output states based upon a set of input variables that are fed into the CPU via the input device, wherein some of the values of the input variables are known and others of the values of the input variables are unknown; and

determining a possibly tentative output decision by comparing the output states with a preselected decision criterion; and

directing the decision explanation module to execute the steps of;

iteratively toggling each input variable that has a known value into an unknown condition, thereby measuring a relative importance of the toggled input variable in terms of a change in the output state;

ranking quantitatively the relative importance of each input variable whose value is known in determining the output decision; and

displaying results of the ranking step on the output device.

8. A computer-implemented process for translating knowledge implicit in a continuous-state feedforward neural network stored in a computer into an explicit set of if-then rules supporting possible conclusions, said computer comprising a CPU and an output device coupled to the CPU, said neural network being stored in said CPU, said CPU comprising an explanation module, said process comprising directing the explanation module to perform the steps of:

determining output states of the neural network based upon a set of input variables, some of whose values are known and others of whose values are unknown;

determining, using the process of claim 10, a minimal subset of possible values of input variables which, if true, would result in reaching a final conclusion condition for the network, where said minimal subset of input values then constitutes a rule for said conclusion condition;

searching input values for a set of all possible rules which support said conclusion condition; and

displaying said rules on said output device.

9. The process of claim 8 wherein, for any continuous-valued input variable, specific input values are chosen to be representative of upper, lower, and middle statistical regions of said variable'"'"'s value range, and those values are examined for possible inclusion in a rule.

11. A computer-implemented process for explaining the reasons for a final conclusion made by a continuous-state feedforward neural network that is stored in a computer, by measuring the contribution to said conclusion made by each input variable having a known value, said computer comprising a CPU, an input device coupled to the CPU, and an output device coupled to the CPU;

said CPU comprising an inference module and a conclusion explanation module;

said neural network having been fed into the CPU via the input device;

said process comprising directing the inference module to perform the steps of;

determining network output states based upon a set of input variables that are fed into the CPU via the input device, some of the values of the input variables being known and others of the values of the input variables being unknown;

determining upper and lower bounds on the output states; and

determining any final conclusion by comparing the bounds with a preselected conclusion criterion; and

directing the conclusion explanation module to perform the steps of;

if a conclusion is possible, iteratively toggling each input variable whose value is known into an unknown condition, thereby measuring a relative importance of the toggled input variable in terms of a change in the upper and lower output bounds;

ranking quantitatively the relative importance of each input variable whose value is known in reaching the final conclusion condition; and

displaying results of the ranking step on the output device.

12. A computer-implemented process for measuring potential influence of an unknown input variable on an output decision of a continuous-state feedforward neural network stored in a computer if the value of the variable becomes known, said computer comprising a CPU and an output device coupled to the CPU, said neural network being stored in the CPU, said CPU comprising an intelligent knowledge acquisition module, said process comprising directing the intelligent knowledge acquisition module to perform the steps of:

determining output states of the neural network based upon a set of input variables, some of whose values are known and others of whose values are unknown;

for each input variable whose value is unknown, iteratively determining the change in the output states as each unknown input variable is toggled to a new known condition within its possible value range, thereby measuring a relative importance of the toggled input variable in terms of a change in the output states;

for each input variable whose value is unknown, ranking quantitatively the relative importance of each said input variable in determining the output decision; and

displaying results of said ranking step on said output device.

13. A computer-implemented process for reducing computation time necessary to determine a modified output state resulting from a modified input state in a continuous-state feedforward neural network stored in a computer, said computer comprising a CPU and an input device coupled to the CPU, said neural network being stored in the CPU, said CPU comprising a module, said process comprising directing the module to perform the steps of:

determining network output states based upon a set of input variables that are fed into the CPU via the input device, wherein said neural network has at least two input processing elements, with each input variable mapped onto at least one input processing element;

changing the value of one of the input variables; and

updating network output states based upon the changed input variable using only those weights associated with connections that lie in paths between input processing elements onto which the changed input variable has been mapped and any output processing element for which an updated state is desired.