Source Code Performance Verification Using Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing software development methods struggle to efficiently verify performance requirements at the source code level, leading to significant rework and inefficiencies, especially for non-compiled programming languages, and existing technologies do not adequately address performance verification during development.
Innovation Solution
A software performance verification system and method using a machine learning model that analyzes source code to predict performance indicators, allowing verification before executable code generation, applicable to various programming languages, including interpreter-type languages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If performance verification is performed using traditional performance analysis tools, then performance can be measured, but verification can only be done at a later stage after executable code generation, requiring significant rework
Solution Approach 1:
The patent performs performance verification in advance by analyzing source code before executable code generation. The performance analysis unit evaluates performance metrics directly from source code, enabling early detection of performance issues and avoiding rework during later development stages.
2Adaptability or versatility
If performance verification is performed at the source code level using traditional methods, then early verification is possible, but it requires compilation and building preparation, which is not applicable to interpreter-type languages
Solution Approach 1:
The patent replaces the mechanical compilation and building process with a machine learning-based performance analysis system. The performance analysis unit uses trained models to directly evaluate source code performance characteristics without requiring compilation, making the verification process applicable to both compiled and interpreter-type languages.
3Measurement precision
If performance verification is performed after executable code generation, then accurate performance measurement is possible, but a large number of preparation man-hours are required for test cases and test data
Solution Approach 1:
The patent creates a virtual performance testing environment by using machine learning models trained on performance data. The performance analysis unit generates predicted performance metrics from source code analysis, eliminating the need for actual executable code and extensive test data preparation while maintaining measurement capability.
Data Source
Figure 1
Figure 2A~2B
Figure 3A~3B
AI summary
A software performance verification system extracts a partial code as part of a code of a program configuring software, generates a feature vector based on the partial code, and generates, as a verification result of the partial code, information based on output obtained from a performance verification model through input of the partial code as a verification target to the performance verification model that is a machine learning model having been trained by use of learning data that includes the feature vector of the partial code for learning and performance information indicative of the performance of software implemented on the basis of the partial code. The software performance verification system generates, for example, a feature vector having metrics values acquired from the partial code as the elements of the feature vector. The performance verification model outputs, for example, the probability of there being a problem with a process to be implemented on the basis of the partial code.