Hyperparameter and network topology selection in network demand forecasting

• US 10,439,891 B2
• Filed: 04/08/2014
• Issued: 10/08/2019
• Est. Priority Date: 04/08/2014
• Status: Active Grant

• Current Assignee: Kyndryl Incorporated
• Original Assignee: International Business Machines Corporation
• Inventors: Baughman, Aaron K., O'Connell, Brian M., Perlitz, Michael, Van Der Stockt, Stefan A. G.
• Primary Examiner(s): Sitiriche, Luis A
• Application Number:
  US14/247,599
• Publication Number:
  US 20150288573A1
• Time in Patent Office: 2,009 Days
• Field of Search: None
• CPC Class Codes: G06N 3/126 Evolutionary algorithms, e.g. genetic algorithms or genetic programming
• CPC Class Codes: H04L 41/145 involving simulating, designing, planning or modelling of a network
• CPC Class Codes: H04L 41/147 for predicting network behaviour

Abstract

Approaches for optimizing network demand forecasting models and network topology using hyperparameter selection are provided. An approach includes defining a pool of features that are usable in models that predict demand of network resources, wherein the pool of features includes at least one historical forecasting feature and at least one event forecasting feature. The approach also includes generating, using a computer device, an optimal model using a subset of features selected from the pool of features. The approach further includes predicting future demand on a network using the optimal model. The approach additionally includes allocating resources in the network based on the predicted future demand.

Claims

1. A method of allocating resources in a networked environment, comprising:

defining a pool of features that are usable in models that predict demand of network resources, wherein the pool of features includes at least one historical forecasting feature which comprises a cyclical demand forecast and at least one event forecasting feature which comprises an analysis of trending topics on a social networking site;

generating, using a computer device, a plurality of models, each model using a unique subset of both historical forecasting features relating to cyclic demand and event forecasting features relating to spike demand on a network, selected from the pool of features;

determining, using the computer device, a relative fitness of each of the plurality of models by testing each of the plurality of models using historical data;

determining, using the computer device, an optimal model from among the plurality of models based upon the determined relative fitness of each of the plurality of models, the optimal model includes a combined forecast curve generated from both the historical forecasting features relating to cyclic demand and event forecasting features relating to spike demand;

predicting future demand on the network using the optimal model; and

allocating resources in the network based on the predicted future demand by allocating additional servers to the network for handling the predicted future demand,

wherein the event forecasting features are used to predict the spike demand based on a number of times at least one of the trending topics is mentioned on the social networking site.

2. The method of claim 1, wherein the defining the pool of features comprises:

defining a set of preprocessing features;

defining a set of historical forecasting features including the at least one historical forecasting feature;

defining a set of post-processing features; and

defining a set of event forecasting features including the at least one event forecasting feature.

3. The method of claim 2, wherein:

the set of preprocessing features includes a plurality of preprocessing functions including a first preprocessing function which calculates seasonality, a second preprocessing function which aligns curves, and a third preprocessing function which inputs values;

the set of historical forecasting features includes a plurality of historical forecasters including an adjusted average forecaster;

the set of post-processing features includes a plurality of post-processing functions including an ensure exists function; and

the set of event forecasting features includes a plurality of event forecasters including a first event forecaster function for simulation;

a second event forecaster function for feature extraction; and

a third event forecaster function for spike prediction.

4. The method of claim 3, wherein the subset of features selected includes:

one of the plurality of preprocessing functions, one of the plurality of historical forecasters, one of the plurality of post-processing functions, and one of the plurality of event forecasters.

5. The method of claim 3, wherein:

the plurality of historical forecasters predict the cyclic demand; and

the plurality of event forecasters predict the spike demand.

6. The method of claim 1, further comprising using an evolutionary algorithm process in the generating the plurality of models.

7. The method of claim 6, wherein the evolutionary algorithm process comprises:

defining a plurality of chromosomes based on combinations of features from the pool of features, wherein a length of a chromosome is a binary vector of length greater than the number of features;

evaluating a fitness of each one of the plurality of chromosomes; and

ranking the plurality of chromosomes based on the evaluated fitness.

8. The method of claim 7, wherein, based on the ranking, a highest ranked one of the plurality of chromosomes includes a combination of features that defines the subset of features used in the optimal model.

9. The method of claim 1, wherein:

the resources in the network are website servers; and

the predicted future demand is predicted traffic on a website.

10. The method of claim 9, wherein the allocating the resources in the network comprises sending at least one instruction to at least one of the website servers to be available for a predicted increase in traffic on the website and wherein the allocating the resources in the network comprises following a predicted demand curve.

11. The method of claim 1, wherein a service provider at least one of creates, maintains, deploys and supports computer infrastructure that includes the computer device.

12. The method of claim 1, wherein steps of claim 1 are provided by a service provider on a subscription, advertising, and/or fee basis.

13. The method of claim 1, wherein, for each of the plurality of models, the relative fitness of the model is determined based on a difference between a demand predicted by the model and an actual historical demand.

14. A computer program product for allocating resources in a networked environment, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions being executable by a computer device to cause the computer device to:

define, by the computer device, a pool of features that includes at least one historical forecasting feature which comprises a cyclical demand forecast and at least one event forecasting feature which comprises an analysis of trending topics on a social networking site, wherein the pool of features excludes redundant features and the features in the pool of features are all predefined;

generate, by the computer device, a plurality of models, each model using a unique subset of both historical forecasting features relating to cyclic demand and event forecasting features relating to spike demand selected from the pool of features;

determine, using the computer device, a relative fitness of each of the plurality of models by testing each of the plurality of models using historical data;

determine, by the computer device, an optimal model from among the plurality of models based upon the determined relative fitness of each of the plurality of models;

predict, by the computer device, future demand on a network using the optimal model; and

allocate, by the computer device, resources in the network based on the predicted future demand by allocating additional servers to the network for handling the predicted future demand,

wherein the event forecasting features are used to predict the spike demand based on a number of times at least one of the trending topics is mentioned on the social networking site.

15. The computer program product of claim 14, wherein the pool of features includes:

a plurality of preprocessing functions;

a plurality of historical forecasters including the at least one historical forecasting feature;

a plurality of post-processing functions; and

a plurality of event forecasters including the at least one event forecasting feature.

16. The computer program product of claim 15, wherein:

the plurality of preprocessing functions are selected using an evolutionary algorithm process;

the plurality of post-processing functions are selected using the evolutionary algorithm process;

the plurality of historical forecasters predict the cyclic demand; and

the plurality of event forecasters predict the spike demand.

17. The computer program product of claim 14, wherein the program instructions cause the computer device to use an evolutionary algorithm process in the generating the plurality of models and the relative fitness of each of the plurality of models is further tested using the historical forecasting features and the event forecasting features.

18. The computer program product of claim 14, wherein:

the resources in the network are website servers;

the predicted future demand is predicted traffic on a website with seasonal traffic demand; and

the allocating the resources comprises sending at least one instruction to at least one of the website servers to be available for a predicted increase in traffic on the website.

19. A system for determining an optimal resource allocation in a networked environment, comprising:

a CPU, a computer readable memory and a computer readable storage medium;

program instructions to define each of a plurality of logical partitions of network resources using a data structure that includes values for a CPU load, disk space, input/output (IO), random access memory (RAM), and network hops that is based on data from a real cloud network;

program instructions to determine the optimal resource allocation using an evolutionary algorithm process and the data structure of each of the plurality of logical partitions; and

program instructions to allocate network resources based on the determined optimal resource allocation,

wherein;

the program instructions are stored on the computer readable storage medium for execution by the CPU via the computer readable memory;

the evolutionary algorithm process includes using historical forecasters that predict cyclic demand and testing using historical data;

the evolutionary algorithm process includes using event forecasters to predict the spike demand based on a number of times a topic is mentioned on a social networking site; and

the evolutionary algorithm process includes defining chromosomes, breeding a next population of chromosomes using crossover and mutation, and a fitness deduction when an expected proportion of probabilistic events types and probabilistic event variances of any required logical partitions of network resources is outside a tolerance value established by a statistical investigation.

20. The system of claim 19, wherein the plurality of logical partitions are defined based on data of the real cloud network and testing using historical data includes comparing the predicted cyclic demand to real cloud historical data, the predicted cyclic demand was predicted from the historical forecasters.

21. The system of claim 19, wherein:

each one of the chromosomes represents a respective cloud with a unique network topology, wherein a length of a chromosome is a binary vector of length equal to all features and functions; and

respective ones of the plurality of logical partitions and type of probabilistic events are genes on the chromosomes.