Adaptive system and method for predicting response times in a service environment

• US 5,444,820 A
• Filed: 12/09/1993
• Issued: 08/22/1995
• Est. Priority Date: 12/09/1993
• Status: Expired due to Fees

• Current Assignee: MarketSpan Corporation (National Grid PLC)
• Original Assignee: Long Island Lighting Company
• Inventors: Tzes, Anthony, Tsotras, Vassilis
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): SHAPIRO, STUART
• Application Number:
  US08/165,020
• Time in Patent Office: 621 Days
• Field of Search: 395/11, 395/22, 395/23, 395/61, 395/900, 395/51
• US Class Current: 706/21
• 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 7/02 using fuzzy logic computing arrangements based on biological models G06N3/00; computing arrangements using knowledge-based models G06N5/00
• CPC Class Codes: Y10S 706/90 Fuzzy logic

Abstract

A hybrid fuzzy logic/neural network prediction system and method is disclosed for predicting response times to service requests to a service provider. Data from a historical database of records including customer requests and weather information are input to the hybrid system. The data is filtered to reject faulty data entries and data not necessarily useful for predicting response times to service requests such as customer comments are eliminated. A backpropagation neural network operating in a supervised learning mode is employed to decrease the effects of the inherent system nonlinearities. The prediction error from the neural network is trained to make predictions within a predetermined error limit. The neural network generates a prediction configuration; i.e. a set of neural network characteristics, for every record per geographical area, time frame, and month. A fuzzy logic classifier is used for further data reliability. A fuzzy logic classifier relying upon the Fuzzy Cell Space Predictor (FCSP) method is employed to improve predicted response times from year to year. The fuzzy logic classifier supervises the overall identification scheme and, for every record, computes a prediction configuration for its corresponding month in the preceding year. The fuzzy logic classifier then computes a prediction estimate for its neighboring months in the preceding year and computes the prediction estimate for the next time frame (i.e. morning and evening). The Center of Gravity method is used to smooth the different prediction estimates to obtain a final predicted response time.

Claims

1. An apparatus for predicting a prospective response time of a service provider to a current service request, the apparatus comprising:

a memory for storing historical data, including previous response times to corresponding previous service requests;

a neural network having a plurality of processing elements, each processing element including memory means for storing a respective weight, said neural network trained by a plurality of test data to modify the weights, said neural network being responsive to a plurality of inputs, including the current month of the current service request and historical data input from the memory for a plurality of months of the previous year for generating a plurality of tentative response time predictions; and

means for classifying the current month of the current service request as similar according to predetermined rules to a first month of the plurality of months of the previous year for generating the prospective response time from a weighting of the plurality of tentative response time predictions in a mapping determined from the current month and the first month.

2. The apparatus as set forth in claim 1, wherein the plurality of processing elements includes an input processing element for receiving a priority number, an output processing element for outputting the prospective response time, and an intermediate processing element for providing a connection between the input processing element and the output processing element.

3. The apparatus as set forth in claim 1, wherein said classifying means classifies the current month using fuzzy logic.

4. The apparatus as set forth in claim 3 wherein the classifying means includes:

a fuzzifier for generating a month similarity value for each of the plurality of months; and

a rule base for applying the predetermined rules to the month similarity values using the fuzzy logic to perform the mapping.

5. The apparatus as set forth in claim 4 wherein the classifying means further includes:

a defuzzifier for generating a non-fuzzy prediction gain from the mapping, and for generating the prospective response time from the weighting of the non-fuzzy prediction gain and the plurality of tentative response time predictions.

6. The apparatus as set forth in claim 1, wherein said neural network further includes:

comparing means for comparing a test response time generated by said neural network responding to the test data to a known response time; and

adjusting means, responsive to said comparing means, for modifying the weights.

7. The apparatus as set forth in claim 6, wherein:

the comparing means generates a root mean squared (RMS) error from said test response time and the known response time; and

the adjusting means uses the RMS error to modify the weights.

8. A method using a neural network for predicting a prospective response time of a service provider to a current service request, the method comprising the steps of:

storing historical data in a memory, the historical data including previous response times to corresponding previous service requests;

training the neural network using the historical data to learn a response time relationship between the previous service requests and the previous response times, the response time relationship being a mapping of the previous service requests to corresponding previous response times;

generating a plurality of weights for the neural network from the learned response time relationship;

generating a plurality of tentative response times from the learned response time relationship by applying a current service request and the historical data from a plurality of months of the previous year to said neural network configured with said plurality of weights;

classifying a current month as similar according to predetermined rules to a first month of the plurality of months;

generating the prospective response time from the tentative response time corresponding to the first month;

storing the generated prospective response time in the memory; and

outputting the prospective response time from the memory.

9. The method as set forth in claim 8, wherein the step of training further comprises the steps of:

a) assigning weights to a plurality of processing elements of the neural network;

b) inputting a known input as a test input to the neural network;

c) generating a test output from the test input using the weights;

d) generating an error from the test output and a known output corresponding to the known input;

e) adjusting the weights using the error; and

f) repeating steps a) to e) until the error is below a predetermined value.