Systems and methods for gaussian process modeling for heterogeneous functions

Patent metadata

Field	Value
Publication number	US-2025094528-A1
Application number	US-202418890651-A
Country	US
Kind code	A1
Filing date	Sep 19, 2024
Priority date	Sep 19, 2023
Publication date	Mar 20, 2025
Grant date	—

Abstract

A system implements a modeling (ClassGP) and an optimization (ClassBO) framework that models heterogeneous functions with knowledge of individual partitions within classes, i.e., heterogeneous functions including non-stationary functions which can be divided into locally stationary functions over partitions of input space with an active stationary function in each partition. The framework constructs a class likelihood for the class by combining log marginal likelihoods associated with each partition of the plurality of partitions within the class. The framework aims to improve analysis of systems that are characterized by a discrete and finite number of “classes” of behaviors.

First claim

What is claimed is:

1 . A system, comprising:
a processor in communication with a memory, the memory including instructions executable by the processor to:
access observation data expressive of operation of an observed system, the observation data including:
a class label associated with an input example of a plurality of input examples, the class label being one of a plurality of class labels; and
a set of closest boundary information for the input example;

determine a plurality of partitions of an input space of the observation data for a class label of the plurality of class labels of the observation data for the observed system using the set of closest boundary information; and
determine a heterogeneous function model for the observed system based on the observation data, the heterogeneous function model for the observed system incorporating a class likelihood function for each class label of the plurality of class labels, the class likelihood function incorporating partition-level log marginal likelihoods for the plurality of partitions of the class label.

2 . The system of claim 1 , the class likelihood function incorporating a summation of partition-level log marginal likelihoods for the plurality of partitions of the class label.

3 . The system of claim 1 , the memory further including instructions executable bye the processor to:
determine class-level hyperparameters of the class likelihood function for the class label by joint optimization of the partition-level log marginal likelihoods for the plurality of partitions of the class label.

4 . The system of claim 1 , each partition of the plurality of partitions of the class label respectively being associated with a partition-level stationary function modeled using a Gaussian process prior distribution having a continuous stationary kernel, the continuous stationary kernel being common to each partition of the plurality of partitions that share the class label.

5 . The system of claim 1 , the observation data including an output example of the observed system that correlates with the input example.

6 . The system of claim 1 , the set of closest boundary information including:
a minimum distance of the input example from a partition of the observation data; and a feature index associated with the minimum distance and the partition.

7 . The system of claim 1 , the memory further including instructions executable by the processor to:
determine the plurality of partitions for the class label of the plurality of class labels of the observation data by construction of a binary classification tree using the set of closest boundary information.

8 . The system of claim 7 , the binary classification tree including a plurality of leaf nodes, each leaf node of the plurality of leaf nodes respectively corresponding to a partition of the plurality of partitions.

9 . The system of claim 7 , the memory further including instructions executable by the processor to:
initialize the binary classification tree using the set of closest boundary information; and refine the binary classification tree using a Metropolis-Hastings sampling scheme conditioned on the observation data.

10 . The system of claim 1 , the memory further including instructions executable by the processor to:
sample a next input example from a distribution of the heterogeneous function model corresponding to the input space for the observed system using an Upper Confidence Bound (UCB) acquisition function.

11 . A method, comprising:
accessing observation data expressive of operation of an observed system, the observation data including:
a class label associated with an input example of a plurality of input examples, the class label being one of a plurality of class labels; and
a set of closest boundary information for the input example;
determining a plurality of partitions of an input space of the observation data for a class label of the plurality of class labels of the observation data for the observed system using the set of closest boundary information; and determining a heterogeneous function model for the observed system based on the observation data, the heterogeneous function model for the observed system incorporating a class likelihood function for each class label of the plurality of class labels, the class likelihood function incorporating partition-level log marginal likelihoods for the plurality of partitions of the class label.

12 . The method of claim 11 , the class likelihood function incorporating a summation of partition-level log marginal likelihoods for the plurality of partitions of the class label.

13 . The method of claim 11 , further comprising:
determining class-level hyperparameters of the class likelihood function for the class label by joint optimization of the partition-level log marginal likelihoods for the plurality of partitions of the class label.

14 . The method of claim 11 , each partition of the plurality of partitions of the class label respectively being associated with a partition-level stationary function modeled using a Gaussian process prior distribution having a continuous stationary kernel, the continuous stationary kernel being common to each partition of the plurality of partitions that share the class label.

15 . The method of claim 11 , the set of closest boundary information including:
a minimum distance of the input example from a partition of the observation data; and a feature index associated with the minimum distance and the partition.

16 . The method of claim 11 , further comprising:
determining the plurality of partitions for the class label of the plurality of class labels of the observation data by construction of a binary classification tree using the set of closest boundary information.

17 . The method of claim 16 , the binary classification tree including a plurality of leaf nodes, each leaf node of the plurality of leaf nodes respectively corresponding to a partition of the plurality of partitions.

18 . The method of claim 16 , further comprising:
initializing the binary classification tree using the set of closest boundary information; and refining the binary classification tree using a Metropolis-Hastings sampling scheme conditioned on the observation data.

19 . The method of claim 11 , further comprising:
sampling a next input example from a probability distribution corresponding to the input space of the heterogeneous function model for the observed system using an Upper Confidence Bound (UCB) acquisition function.

20 . A non-transitory computer readable medium including instructions encoded thereon that are executable by a processor to:
access observation data expressive of operation of an observed system, the observation data including:
a class label associated with an input example of a plurality of input examples, the class label being one of a plurality of class labels; and
a set of closest boundary information for the input example;
determine a plurality of partitions of an input space of the observation data for a class label of the plurality of class labels of the observation data for the observed system using the set of closest boundary information; and determine a heterogeneous function model for the observed system based on the observation data, the heterogeneous function model for the observed system incorporating a class likelihood function for each class label of the plurality of class labels, the class likelihood function incorporating partitio

Assignees

• Malu Mohit
• Pedrielli Giulia
• Dasarathy Gautam
• Spanias Andreas
• Univ Arizona State

Inventors

• Malu Mohit
• Pedrielli Giulia
• Dasarathy Gautam
• Spanias Andreas

Classifications

• G06F17/13 Primary

  Differential equations (using digital differential analysers G06F7/64) · CPC title