Gaussian Process Prediction Method for Time-Series Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional generative model methods for time-series prediction face high calculation costs and low prediction accuracy, especially for complex data, due to the need for repeated RNN calculations and Monte Carlo simulations.
Innovation Solution
A prediction method that optimizes parameters of a Gaussian process and a neural network to calculate a prediction distribution for future observation values, using a series of observation values and covariates, thereby reducing calculation costs while improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional generative model methods (RNN + Monte Carlo simulation) are used for time-series prediction, then prediction accuracy can be maintained, but calculation cost becomes enormous especially as prediction period increases
Solution Approach 1:
The patent transforms the target values through a nonlinear function (first function) so that the transformed values follow a Gaussian process. This parameter transformation allows the use of efficient Gaussian process prediction methods instead of computationally expensive RNN + Monte Carlo simulation approaches, thereby reducing calculation cost while maintaining prediction accuracy
Solution Approach 2:
The patent replaces the mechanical computation process of RNN + Monte Carlo simulation with a mathematical transformation approach based on Gaussian processes. By substituting the iterative simulation mechanism with a closed-form Gaussian process solution, the calculation cost is dramatically reduced while preserving prediction accuracy
2Adaptability or versatility
If conventional generative model methods are used for complex time-series data, then prediction can be performed, but prediction accuracy becomes low
Solution Approach 1:
The patent applies a nonlinear transformation (first function) to the target values to make them conform to Gaussian process assumptions. This parameter transformation enables the Gaussian process model to accurately capture complex temporal dependencies and patterns in the data, improving prediction accuracy for complex time-series while maintaining adaptability
3Duration of action of moving object
If the number of prediction steps is increased, then longer prediction period is achieved, but calculation cost increases proportionally
Solution Approach 1:
The patent performs preliminary transformation of the target values through the first function during the training phase, establishing the Gaussian process parameters in advance. This preliminary action allows for efficient prediction at any future time point without requiring repeated RNN calculations and Monte Carlo simulations, thus enabling long prediction periods with controlled calculation cost
Solution Approach 2:
By transforming the target values to follow a Gaussian process and optimizing the kernel function parameters, the patent enables efficient computation of predictions for multiple future time points. The Gaussian process framework allows for analytical solutions that scale better with prediction horizon compared to iterative RNN + Monte Carlo approaches
Data Source
AI summary
A prediction method executed by a computer including a memory and a processor, the method includes: optimizing a parameter of a second function that outputs parameters of a first function from covariates, and optimizing a parameter of a kernel function of a Gaussian process, by using a series of observation values observed in a past and a series of the covariates observed simultaneously with the observation values, wherein values obtained by non-linearly transforming the observation values by the first function follow the Gaussian process; and calculating a prediction distribution of observation values in a period in future to be predicted by using the second function and the kernel function having parameters optimized in the optimizing, and a series of covariates in the period.