Computer system and method for distributing execution of a predictive model

Patent metadata

Field	Value
Publication number	US-10878385-B2
Application number	US-201715599360-A
Country	US
Kind code	B2
Filing date	May 18, 2017
Priority date	Jun 19, 2015
Publication date	Dec 29, 2020
Grant date	Dec 29, 2020

Abstract

Disclosed herein are systems, devices, and methods related to assets and predictive models and corresponding workflows that are related to the operation of assets. In particular, examples involve assets configured to receive and locally execute predictive models, locally individualize predictive models, and/or locally execute workflows or portions thereof.

First claim

The invention claimed is:

1. A computing system comprising:
a network interface;
a memory;
at least one processor;
a non-transitory computer-readable medium; and
program instructions stored on the non-transitory computer-readable medium that are executable by the at least one processor to cause the computing system to:
select a given predictive model that is configured to predict a problem at an asset;
define an approximated version of the given predictive model by comparing, for each of plurality of subsets of input values in a set of training data, (a) a loss associated with executing an approximated version of the given predictive model locally at an asset for a subset of input values and (b) a cost associated with executing the given predictive model remotely at the computing system for the subset of input values, wherein the approximated version of the given predicted model comprises a set of approximation functions and corresponding base regions, and wherein each respective base region represents a subset of input values for which a loss associated with executing the corresponding approximation function locally at the asset is lower than a cost associated with executing the given predictive model remotely at the computing system; and
deploy the approximated version of the given predictive model for local execution by the asset.

2. The computing system of claim 1 , wherein the program instructions that cause the computing system to define the approximated version of the given predictive model comprise program instructions that cause the computing system to:
determine a candidate base region;
determine a transmission cost for the candidate base region;
determine an approximation function corresponding to the candidate base region;
determine a loss associated with errors of the approximation function of the candidate base region; and
incorporate the approximation function and the candidate base region into the approximated version of the given predictive model if the approximated version of the given predictive model has a loss associated with errors of the approximation function that is lower than the transmission cost for the candidate base region.

3. The computing system of claim 1 , further comprising program instructions stored on the non-transitory computer-readable medium that are executable by the at least one processor to cause the computing system to:
receive, from the asset, input data that is not within any of the base regions; and
for at least the received input data, execute the given predictive model at the computing system.

4. The computing system of claim 1 , wherein the program instructions that cause the computing system to define the approximated version of the given predictive model comprise program instructions that cause the computing system to:
determine, based on the base region, a candidate base region, wherein the candidate base region is a superset of the base region;
determine a loss associated with errors of the approximation function for the candidate base region; and
expand the base region to encompass the candidate base region if the loss of approximation function for the candidate base region is less than or equal to the loss of the approximation function for the base region.

5. The computing system of claim 1 , wherein the program instructions that cause the computing system to define the approximated version of the given predictive model comprise program instructions that cause the computing system to:
determine, based on one of the base regions, a candidate expanded base region, wherein the candidate expanded base region is a superset of the base region;
determine a candidate approximation function for the expanded base region;
determine a loss associated with errors of the candidate approximation function for the expanded base region; and
if loss cost of the candidate approximation function for the candidate expanded base region is less than or equal to the loss of the approximation function over the base region, and if the loss of the candidate approximation function for the candidate expanded region is less than or equal to the loss of the approximation function for the base region:
expand the base region to encompass the candidate expanded base region in the approximated version of the given predictive model; and
replace the approximation function with the candidate approximation function in the approximated version of the given predictive model.

6. The computing system of claim 5 , wherein the program instructions that cause the computing system to define the approximated version of the given predictive model comprise program instructions that cause the computing system to:
attempt to expand the base region using a binary search of candidate expanded base regions that are sorted based on sizes of the candidate expanded base regions.

7. The computing system of claim 1 , wherein the program instructions that cause the computing system to define the approximated version of the given predictive model comprise program instructions that cause the computing system to:
generate multiple approximated version of the given predictive model; and
select for deployment to the asset, based on a loss function, whichever one of the multiple approximated versions of the given predictive model has a lowest loss.

8. A method comprising:
selecting a given predictive model that is configured to predict a problem at an asset;
defining an approximated version of the given predictive model by comparing, for each of plurality of subsets of input values in a set of training data, (a) a loss associated with executing an approximated version of the given predictive model locally at an asset for a subset of input values and (b) a cost associated with executing the given predictive model remotely at a computing system for the subset of input values, wherein the approximated version of the given predicted model comprises a set of approximation functions and corresponding base regions, and wherein each respective base region represents a subset of input values for which a loss associated with executing the corresponding approximation function locally at the asset is lower than a cost associated with executing the given predictive model remotely at the computing system; and
deploying the approximated version of the given predictive model for local execution by the asset.

9. The method of claim 8 , wherein defining the approximated version of the given predictive model comprises:
determining a candidate base region;
determining a transmission cost for the candidate base region;
determining an approximation function corresponding to the candidate base region;
determining a loss associated with errors of the approximation function of the candidate base region; and
incorporating the approximation function and the candidate base region into the approximated version of the given predictive model if the approximated version of the given predictive model has a loss associated with errors of the approximation function that is lower than the transmission cost for the candidate base region.

10. The method of claim 8 , further comprising:
receiving, from the asset, input data that is not within any of the base regions; and
for at least the received input data, executing the given predictive model at the computing system.

11. The method of claim 8 , wherein defining the approximated version of the given predictive model comprises:
determining, based on the base region, a candidate base region, wherein the candidate base region is a superset of the base region;
determining a loss associated with errors of the approximation function for the candi

Assignees

• Uptake Tech Inc

Inventors

• Horrell Michael
• Mcelhinney Adam

Classifications

• G06F11/3058

  Monitoring arrangements for monitoring environmental properties or parameters of the computing system or of the computing system component, e.g. monitoring of power, currents, temperature, humidity, position, vibrations (thermal management in cooling arrangements of a computing system G06F1/206) · CPC title

• G06Q10/0635

  Risk analysis of enterprise or organisation activities · CPC title

• G06F9/48

  Program initiating; Program switching, e.g. by interrupt · CPC title

• G06F9/4856

  resumption being on a different machine, e.g. task migration, virtual machine migration (G06F9/5088 takes precedence) · CPC title

• G06F11/008

  Reliability or availability analysis · CPC title