Predictive Engine for Multivariate Categorical Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for optimizing multiple categorical outputs are computationally complex and inefficient, failing to provide accurate confidence estimates and often do not converge to an optimized solution, especially when dealing with correlated outputs and large datasets.
Innovation Solution
A predictive engine using Bayesian optimization techniques with latent continuous variables to model and optimize multivariate categorical responses, incorporating dependency measurement and coefficient estimation through joint posterior density distributions, enabling efficient computation and accurate probability assessments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Bayesian optimization methods are applied to multivariate outcome problems, then confidence estimates and correlation structure are provided, but computational complexity and cost increase significantly
Solution Approach 1:
The patent segments the multivariate optimization problem into multiple univariate response surface models, each modeling a single output variable independently. This segmentation allows traditional optimization methods to be applied to each univariate model separately, avoiding the computational complexity of multivariate Bayesian optimization while still providing confidence estimates through the ensemble of univariate models.
Solution Approach 2:
The patent introduces desirability functions as intermediary transformations that convert multiple output variables into a single composite desirability score. This intermediary representation enables the use of traditional univariate optimization techniques while indirectly addressing the multivariate optimization problem, reducing computational complexity while maintaining reliability.
2Productivity
If traditional response surface optimization is used, then computational efficiency is maintained, but accuracy and certainty estimates of optimized settings are lost
Solution Approach 1:
The patent implements feedback by iteratively refining the univariate response surface models and their associated desirability functions based on observed data. Each iteration provides updated confidence estimates and adjusts the models to better reflect the underlying relationships, maintaining computational efficiency while progressively improving measurement precision through data-driven refinement.
Data Source
AI summary
A predictive engine includes an optimizer and a predictor. The optimizer is configured to receive an observed dataset having inputs and multivariate responses and determine latent response variables based on the predictive inputs and the multivariate responses. The optimizer is further configured to select latent response variables, measure dependencies between multivariate responses, estimate coefficients that relate the input predictors to determined latent response variables, and correlate dependencies and coefficients with the latent response variables. The predictor is configured to generate a predictive distribution of probabilities for the latent variables, map the probabilities to multivariate responses, generate a predictive distribution of probabilities for the multivariate responses, and determine at least one optimized input from the multivariate responses.