Genetic Classification Program Fitness via Identification Margin Distribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for automatically generating class classification programs using genetic programming often fail to produce programs with high general versatility, particularly when the number of learning images is limited, leading to insufficient classification accuracy due to minimal differences in correct answer rates and identification margin values.
Innovation Solution
An information processing apparatus and method that execute genetic processing on class classification programs to calculate a fitness metric based on both the correct answer rate and the separability between classes, using the distance from an identification boundary, to automatically generate a class classification program with high general versatility by adjusting the fitness calculation equation to prioritize separability and correct answer rate.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If genetic programming is used to automatically generate class classification programs, then automation is improved, but classification accuracy deteriorates when the number of learning images is limited
Solution Approach 1:
The patent changes the evaluation parameter from simple correct answer rate to a composite fitness function that includes correct answer rate, identification margin, and distribution width. This parameter transformation allows the genetic programming to optimize for both accuracy and generalization capability simultaneously, resolving the contradiction between automation and classification accuracy.
Solution Approach 2:
The patent introduces a new dimension of evaluation by considering the distribution width of identification margins in addition to correct answer rate. This dimensional expansion of the evaluation space allows the system to distinguish between programs that achieve high accuracy through overfitting versus those that achieve it through genuine generalization capability.
2Device complexity
If the fitness calculation focuses only on correct answer rate, then simplicity is improved, but general versatility deteriorates
Solution Approach 1:
The patent transforms the fitness calculation from a single-parameter (correct answer rate) evaluation to a multi-parameter evaluation system that incorporates correct answer rate, identification margin, and distribution width. This parameter expansion enables the system to evaluate and select programs with better general versatility without excessive complexity.
Solution Approach 2:
The patent implements a feedback mechanism where the distribution width of identification margins is calculated and fed back into the fitness function. This feedback allows the genetic programming to iteratively improve programs by selecting those that not only achieve high accuracy but also maintain wide distribution of identification margins, indicating better generalization.
3Quantity of substance
If the number of learning images is limited, then data quantity is reduced, but classification accuracy deteriorates due to insufficient separability
Solution Approach 1:
The patent changes the evaluation focus from merely counting correct answers to analyzing the distribution characteristics of identification margins. By incorporating distribution width into the fitness function, the system can effectively utilize limited learning images to train programs that generalize well, overcoming the data quantity limitation.
Solution Approach 2:
The patent performs preliminary analysis of identification margin distributions during the genetic programming process, selecting parent programs based on their distribution characteristics before generating offspring. This preliminary action ensures that programs with better generalization potential are preferentially selected and combined, maximizing the utility of limited learning data.
Data Source
AI summary
An information processing apparatus includes: a memory; and a processor configured to: execute genetic processing on class classification programs of a set of class classification programs; acquire a distance between each learning data and an identification boundary regarding evaluation identifiers created by using first class learning data belonging to a first class and second class learning data belonging to a second class according to the class classification program; calculate a statistic amount of a distribution of a distance to the identification boundary for each of the first class learning data and the second class learning data; define a fitness calculation equation based on the statistic amounts of the first class learning data and the second class learning data; calculate the fitness of the class classification program; and determine whether or not to replace the class classification program with one of class classification programs of the set depending on the fitness.