Patent US936654A successfully refreshed from Google Patents.
Patent EP0468229A2 successfully refreshed from Google Patents.
EP0468229A2
Expired / Lapsed
EP
⏱️ 0.0 years remaining
A neural network with expert system functionality
Assignee: HNC Inc
•
Filing: 1991-07-01
•
Granted: —
•
Expires: 2011-07-01
Abstract
Abstract A method is disclosed for performing a variety of expert system functions on any trained 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.
Inventors
- 👤 Curt A. Levey
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-11-05 08:47
CPC Classifications
G; G06; G06N; G06N5/00; G06N5/02; G06N5/022; G06N3/00; G06N3/02; G06N3/04; G06N3/042; G06N5/04; G06N5/045
Claims Analysis
Total Claims: 17
Claim 1
Independent Claim
173 words
A method of operating an expert system based on a trained continuous-state feedforward neural network, 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, the method comprising the steps of: determining a decision for each output variable based on the states of the underlying PEs; determining a conclusion for each output variable when, as inputs to the network become known, the decision becomes irreversible; computing confidence measures for each decision; explaining each decision by computing the contribution made toward the decision by each input variable with a known value; explaining each conclusion by computing the contribution made toward reaching the conclusion by each input variable with a known value; determining the potential influence on a decision of each input variable with an unknown value; and converting the knowledge implicit in the neural network into an explicit set of if-then rules.
Claim 2
Independent Claim
173 words
A method of operating an expert system based on a trained continuous-state feedforward neural network, 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, the method comprising the steps of: determining a decision for each output variable based on the states of the underlying PEs; determining a conclusion for each output variable when, as inputs to the network become known, the decision becomes irreversible; computing confidence measures for each decision; explaining each decision by computing the contribution made toward the decision by each input variable with a known value; explaining each conclusion by computing the contribution made toward reaching the conclusion by each input variable with a known value; determining the potential influence on a decision of each input variable with an unknown value; and converting the knowledge implicit in the neural network into an explicit set of if-then rules.
Claim 3
Independent Claim
111 words
A method for computing high and low bounds on the states of hidden and output PEs a feedforward neural network, wherein these bounds serve as limits on the possible range of each such state, given known input states and any combination of possible states for some or all of the input PEs whose states are currently unknown, the method comprising the steps of: determining the upper and lower bounds on the values of continuous input variables, using the input variable distribution statistics; and computing the bounds on the states of the output PEs recursively from the bounds on the input variables and the bounds on the states of the hidden PEs.
Claim 4
Independent Claim
111 words
A method for computing high and low bounds on the states of hidden and output PEs a feedforward neural network, wherein these bounds serve as limits on the possible range of each such state, given known input states and any combination of possible states for some or all of the input PEs whose states are currently unknown, the method comprising the steps of: determining the upper and lower bounds on the values of continuous input variables, using the input variable distribution statistics; and computing the bounds on the states of the output PEs recursively from the bounds on the input variables and the bounds on the states of the hidden PEs.
Claim 5
Dependent Claim
85 words
A method according to claim 2 for determining when a conclusion can be made regarding the value of an output variable in a feedforward neural network with continuous-state PEs, despite some number of unknown input states, the method comprising the steps of: determining the high and low bounds on the output states; using the high and low bounds on the output states to determine if the tentative decision for the output variable can ever change; and producing a conclusion if the tentative decision cannot change.
Claim 6
Independent Claim
69 words
A method for computing a confidence measure of decisiveness for a decision corresponding to the value of an output variable in a feedforward neural network, the method comprising the steps of: determining the network output states given a set of known input values; and comparing the output states with the selected threshold or other decision criteria and quantitatively measuring the closeness of the output states to the decision criteria.
Claim 7
Independent Claim
69 words
A method for computing a confidence measure of decisiveness for a decision corresponding to the value of an output variable in a feedforward neural network, the method comprising the steps of: determining the network output states given a set of known input values; and comparing the output states with the selected threshold or other decision criteria and quantitatively measuring the closeness of the output states to the decision criteria.
Claim 8
Dependent Claim
81 words
A method according to claim 2 for computing a confidence measure of certainty for a decision corresponding to the value of an output variable in a feedforward neural network, the method comprising the steps of: determining the bounds on the states of the output PEs; comparing the output bounds with the selected threshold or other decision criteria and quantitatively measuring the closeness of the bounds to the decision criteria to determine the probability that the output decision will not be reversed.
Claim 9
Independent Claim
91 words
A method of explaining the reasons for a neural network output decision, the method comprising the steps of: determining the network output states on a set of known input values; comparing the output states with the decision criteria and determining an output decision; sequentially converting each known input value to an unknown condition; for each input value which is converted from known to unknown, measuring the change in the output state relative to the decision criteria; and ranking quantitatively the importance of each known input value in determining the output decision.
Claim 10
Independent Claim
91 words
A method of explaining the reasons for a neural network output decision, the method comprising the steps of: determining the network output states on a set of known input values; comparing the output states with the decision criteria and determining an output decision; sequentially converting each known input value to an unknown condition; for each input value which is converted from known to unknown, measuring the change in the output state relative to the decision criteria; and ranking quantitatively the importance of each known input value in determining the output decision.
Claim 11
Dependent Claim
96 words
A method of explanation according to claim 3 for measuring the contribution made toward reaching the conclusion condition by each input variable with a known value in a feedforward neural network, the method comprising the steps of: determining the upper and lower output state bounds; sequentially converting each known input value to the unknown condition; for each input value which is converted from known to unknown, measuring the change in the upper and lower output bounds relative to the decision criteria; and ranking quantitatively the importance of each known input value in reaching the conclusion condition.
Claim 12
Dependent Claim
66 words
A method according to claim 3 for computing a minimal subset of current input values that is sufficient to support the conclusion condition in a continuous-state feedforward neural network, the method comprising the steps of: evaluating the relative changes in the output bounds for each known input value; and selecting the minimum set of the most important inputs which are sufficient to reach the conclusion condition.
Claim 13
Independent Claim
79 words
A method of measuring the potential influence an unknown input variable may have on the decision corresponding to the output of a feedforward neural network if the value of the variable becomes known, the method comprising the steps of: for each unknown input value, sequentially determining the change in the output state as each input is varied over its possible value range; and comparing quantitatively the changes in output states relative to the decision criteria for each unknown input.
Claim 14
Independent Claim
79 words
A method of measuring the potential influence an unknown input variable may have on the decision corresponding to the output of a feedforward neural network if the value of the variable becomes known, the method comprising the steps of: for each unknown input value, sequentially determining the change in the output state as each input is varied over its possible value range; and comparing quantitatively the changes in output states relative to the decision criteria for each unknown input.
Claim 15
Dependent Claim
88 words
A method according to claim 3 for using the conclusion generating process to translate the knowledge implicit in a continuous-state feedforward neural network into an explicit set of if-then rules supporting possible conclusions, the method comprising the steps of: determining a minimum set of input values which will result in reaching the conclusion condition, where the set of input values then constitute a rule for the particular conclusion; and searching input values for the set of all possible rules which support any conclusion for a particular output variable.
Claim 16
Dependent Claim
51 words
A method of using the values of continuous input variables in the rules according to claim 10 wherein specific values are chosen to be representative of the upper, lower, and middle statistical regions of the variable's value range, and those values are examined for possible inclusion in one or more rules.
Claim 17
Dependent Claim
85 words
A method according to claim 6 for using a subset of the weights of a feedforward neural network to compute the modified output states that result from an actual change or an assumed change in input values, the method comprising the additional steps of using the weights associated with connections that lie in a path between one of the input PEs whose state is being changed and one of the output PEs for which a modified state is being computed to compute the modified states.
References Cited (Backward Citations)
No citation records indexed.
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.