Model Selection Device for Mixed Distributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing model selection methods struggle with high-speed selection for complex mixed distribution models, especially when components are dependent and the number of component candidates increases exponentially, making it impractical to calculate information criteria for all model candidates.
Innovation Solution
A model selection device and method that optimizes the expected information criterion for complete data with respect to a hidden variable post-event distribution, allowing for high-speed model selection by optimizing a pair of a model and a parameter that satisfies a predetermined condition, even when components are dependent and the number of candidates is large.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If model selection methods calculate information criteria for all model candidates to ensure accurate model selection, then model selection accuracy is improved, but computational complexity increases exponentially
Solution Approach 1:
The patent segments the model selection process into two stages: first optimizing parameters for each candidate model using EM algorithm, then selecting the optimal model based on information criterion. This segmentation allows efficient parameter optimization independent of model structure complexity, reducing overall computational burden while maintaining selection accuracy.
Solution Approach 2:
The patent performs preliminary parameter optimization for each candidate model before evaluating the information criterion. By pre-optimizing parameters using EM algorithm, the method ensures that each candidate model is evaluated at its best possible performance, avoiding redundant calculations during the selection phase and reducing total computational complexity.
2Adaptability or versatility
If the number of model candidates is increased to cover more complex mixed distribution models, then model selection versatility is improved, but calculation time increases exponentially
Solution Approach 1:
The patent separates parameter optimization from model selection, allowing the same efficient EM algorithm to be applied regardless of the number of candidate models. This segmentation enables the method to handle a large number of candidates (including complex mixed distributions) without proportionally increasing computational time, as each candidate is processed independently through the same optimized workflow.
Solution Approach 2:
The patent changes the parameter optimization approach by using EM algorithm that converges to optimal parameters regardless of the specific model structure. This parameter optimization method is universally applicable to different mixed distribution models (Gaussian, Poisson, etc.), allowing the system to handle diverse and complex models efficiently without requiring model-specific optimization routines.
3Reliability
If parameter optimization is performed for each candidate model to ensure accurate information criterion calculation, then model selection reliability is improved, but processing speed decreases
Solution Approach 1:
The patent replaces brute-force exhaustive search with the EM algorithm for parameter optimization. The EM algorithm provides a systematic, iterative approach that converges to optimal parameters efficiently, substituting the mechanical trial-and-error approach with a mathematically grounded optimization method that maintains reliability while improving processing speed.
Solution Approach 2:
The patent performs parameter optimization as a preliminary step before information criterion evaluation. By completing parameter optimization beforehand, the method ensures reliable model selection while avoiding repeated optimization during the selection process, thereby maintaining processing speed. The pre-optimized parameters are then reused for evaluating different model candidates.
Data Source
AI summary
The model selection device comprises a model optimization unit which optimizes a model for a mixed distribution, wherein related to an information criterion of complete data, with respect to a hidden variable post-event distribution of the complete data, the model optimization unit optimizes an expected information criterion of the complete data for a pair of a model and a parameter of a component which satisfies a predetermined condition.