# US 2005/0097207 A1 · System and method of predicting future behavior of a battery of end-to-end probes to anticipate and prevent computer network performance degradation

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2005/0097207 A1
- Document ID: US20050097207A1
- Status: Published application
- Kind: A1
- Application number: 10/697,273
- Filed: 2003-10-31
- Published: 2005-05-05
- Earliest claimed: 2003-10-31
- Assignees: International Business Machines Corporation
- Inventors: Ilya Gluhovsky; Alan Hoffman; Herbert Lee; Emmanuel Yashchin
- Web page: https://patentgrep.com/patent/US20050097207A1

## CPC

- H04L 43/12: Network monitoring probes
- H04L 41/0681: Configuration of triggering conditions
- H04L 41/142: using statistical or mathematical methods
- H04L 41/147: for predicting network behaviour
- H04L 41/149: for prediction of maintenance
- H04L 41/5025: by proactively reacting to service quality change, e.g. by reconfiguration after service quality degradation or upgrade
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

A diagnostic system in which, at every point in time, a forecast is made of the future response times of each EPP (End-to-end Probe Platform) probe in a battery of probes. Thresholds are established in terms of the distribution of future EPP values. The theory of Generalized Additive Models is used to build a predictive model based on a combination of a) data normally generated by network nodes, b) results of a battery of probes and c) profile curves reflecting expected response times (i.e. based on recent history) corresponding to this battery for various times of day, days of week, month of year, etc. The model is pre-computed, and does not have to be dynamically adjusted. The model produces, at regular intervals, forecasts for outcomes of various EPP probes for various horizons of interest; also, it produces thresholds for the respective forecasts based on a number of factors, including acceptable rate of false alarms, forecast variance and EPP values that are expected based on the recorded history. The system is capable of maintaining a pre-specified low rate of false alarms that could otherwise cause a substantial disturbance in network operation.

## Claims

### Claim 1

A method for using future behavior of a battery of probes to forecast and diagnose network performance, comprising the steps of:
consolidating normal network data and data returned from a battery of EPP probes into a single model matrix, a probe data set in said returned data including at least a time of said probe and a user performance measure;
predicting a future value of said user performance measure from predictor variables in said single model matrix, values of said predictor variables being lagged from said probe time by a prediction horizon;
generating an alarm signal if said predicted future value of the performance measure is above a threshold value;
observing an actual value of said performance measure at the predicted time, said alarm signal being FALSE if the actual value is below a lower alarm threshold and an alarm signal being MISSED if no alarm signal was generated and the actual value is above an upper alarm threshold; and
setting by trial and error said threshold value and said prediction horizon to achieve a rate of said false alarms below a specified limit.

### Claim 2 (depends on 1)

The method of claim 1, wherein said predicting step further comprises the steps of:
reducing a dimensionality of said single model matrix by performing stepwise linear regression, at each step adding a most significant network node and deleting a least significant network node, a standard F test providing a significance statistic, until no addition or deletion is significant;
fitting functions of said predictor variables to an additive model; and
generating from said additive model an estimated mean of said future value of said user performance measure of said EPP probe, an estimated variance of said estimated mean, and for each predictor variable an estimated variance of an estimated contribution to said future value of said predictor variable.

### Claim 3 (depends on 1)

The method of claim 1, further comprising the step of deriving a control scheme for said EPP probe by selecting coefficients for a formula for said alarm threshold, said alarm threshold formula including a constant term, a seasonality term and a term containing an estimated standard deviation of said future value of said user performance measure, wherein a rate of said missed alarms in said trial and error setting step is determined to be acceptable.

### Claim 4 (depends on 1)

The method of claim 1, wherein said model matrix includes for each EPP probe said user performance measure is a response time Y;
said time of said probe is a time t when said response time Y was observed;
said prediction horizon has a value h; and
wherein for each of said predictor variables said lagged value is taken at, or most recently prior to, time t-h.

### Claim 5 (depends on 1)

The method of claim 1, further comprising the steps of:
obtaining a list of models, one for each said linear regression step, a model for a last step providing a selected subset of predictor variables;
using a partial stepwise additive procedure to further reduce said dimensionality, said dimensionality being the number of a plurality of said network nodes; and
determining a scale of EPP probe values, said scale being ordered by predictive power.

### Claim 6 (depends on 1)

The method of claim 1, further comprising the step of determining interaction terms for said additive model, said interaction terms being a function of more than one predictor variable.

### Claim 7 (depends on 3)

The method of claim 3, wherein said step of deriving a control scheme uses profile curves, said profile curves reflecting expected values of said user performance measure based on normally observed regularities in variation of values of said user performance measure.

### Claim 8 (depends on 7)

The method of claim 7, wherein said normally observed regularities include variations in a weekly cycle and variations in a daily cycle.

### Claim 9 (depends on 1)

The method of claim 1, further comprising the step of presenting a graphical summary of unfavorable predictors in response to triggering of said alarm threshold.

### Claim 10 (depends on 2)

The method of claim 2, further comprising the steps of
determining whether said additive model is adequate and triggering a model alarm if said additive model is determined not to be adequate; and
identifying elements of said additive model whose characteristics triggered said model alarm, said identifying step being initiated by said model alarm.

### Claim 11 (depends on 10)

The method of claim 10, further comprising the steps of:
selecting a new predictive model from a battery of recovery models; and
building said new predictive model over time as data is accumulated for said predictor variables.

### Claim 12 (depends on 1)

The method of claim 1, wherein a predictive model is developed for a plurality of prediction horizons, one model at a time.

### Claim 13

A system for using future behavior of a battery of probes to forecast and diagnose network performance, comprising:
a single model matrix comprising normal network data and responses to a battery of probes;
profile curves corresponding to said battery, said profile curves reflecting expected response times based on recent history of said probes;
an additive model that predicts the behavior of a function of said probes at a future time, said function being defined in terms of predictor variables in said model matrix, said function being selected by advanced statistical methods, and said additive model being applied cyclically to reduce a variable space of said predictor variables in said single model matrix;
alarm thresholds based on deviation between said profile curves and behavior predicted by said additive model, said alarm thresholds being selected to assure a predictably low rate of false alarms; and
a graphical summary of unfavorable predictors in response to triggering of said alarm threshold.

### Claim 14 (depends on 13)

A system as in claim 13, further comprising:
a monitor for determining whether said additive model is adequate, said monitor triggering a model alarm if said additive model is determined not to be adequate; and
a diagnostic for identifying elements of said additive model whose characteristics triggered said model alarm, said diagnostic being initiated by said model alarm.

### Claim 15 (depends on 13)

A system as in claim 13, wherein said cyclic application of said model uses an F statistic to control addition and deletion of variables in said variable space.

### Claim 16 (depends on 13)

A system as in claim 13, wherein said cyclic application of said model uses a subset of variables determined by a stepwise linear regression (SLR).

### Claim 17

A computer implemented system for using future behavior of a battery of probes to forecast and diagnose network performance, comprising:
computer code for consolidating normal network data and data returned from a battery of EPP probes into a single model matrix, a probe data set in said returned data including at least a time of said probe and a user performance measure;
computer code for predicting a future value of said user performance measure from predictor variables in said single model matrix, values of said predictor variables being lagged from said probe time by a prediction horizon;
computer code for generating an alarm signal if said predicted future value of the performance measure is above a threshold value;
computer code for observing an actual value of said performance measure at the predicted time, said alarm signal being FALSE if the actual value is below a lower alarm threshold and an alarm signal being MISSED if no alarm signal was generated and the actual value is above an upper alarm threshold; and
computer code for setting by trial and error said threshold value and said prediction horizon to achieve a rate of said false alarms below a specified limit.

### Claim 18 (depends on 17)

The computer implemented system of claim 17, wherein said computer code for predicting further comprises:
computer code for reducing a dimensionality of said single model matrix by performing stepwise linear regression, at each step adding a most significant network node and deleting a least significant network node, a standard F test providing a significance statistic, until no addition or deletion is significant;
computer code for fitting functions of said predictor variables to an additive model; and
computer code for generating from said additive model an estimated mean of said future value of said user performance measure of said EPP probe, an estimated variance of said estimated mean, and for each predictor variable an estimated variance of an estimated contribution to said future value of said predictor variable.

### Claim 19 (depends on 17)

The computer implemented system of claim 17, further comprising computer code for deriving a control scheme for said EPP probe by selecting coefficients for a formula for said alarm threshold, said alarm threshold formula including a constant term, a seasonality term and a term containing an estimated standard deviation of said future value of said user performance measure, wherein a rate of said missed alarms in said trial and error setting step is determined to be acceptable.

### Claim 20 (depends on 18)

The computer implemented system of claim 18, further comprising:
computer code for determining whether said additive model is adequate and for triggering a model alarm if said additive model is determined not to be adequate; and
computer code for identifying elements of said additive model whose characteristics triggered said model alarm, said identifying step being initiated by said model alarm.
