Optimization Models From Verification Code Using Transformer Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The scarcity of optimization experts and the inefficiencies in generating decision optimization specifications lead to sub-optimal solutions for optimization issues, which are often resolved heuristically or with custom tools, failing to effectively utilize resources and time.
Innovation Solution
A computer-implemented method using a transformer-based deep learning network to generate and fine-tune optimization solutions by converting optimization solution verification programs into loss functions, incorporating constraints and objectives, and training a sequence generation model with labeled random inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual optimization model generation is used, then solution verification accuracy is improved, but productivity and time efficiency deteriorate
Solution Approach 1:
The patent uses large language models to automatically generate optimization model specifications by learning from and copying patterns in existing verification programs and solution data, replacing manual model generation while maintaining accuracy through trained AI systems
Solution Approach 2:
The patent replaces manual mechanical processes of optimization model creation with automated AI-based systems that use machine learning to generate and verify optimization models, significantly improving productivity while maintaining solution accuracy
2Adaptability or versatility
If custom tools and heuristic methods are used, then implementation flexibility is improved, but solution optimality and resource utilization deteriorate
Solution Approach 1:
The patent automatically adjusts and optimizes parameters in optimization models using AI-generated specifications, ensuring optimal resource utilization and solution quality while maintaining the flexibility to adapt to different problem types through programmable parameters
3Measurement precision
If expert manual generation is used, then model accuracy is improved, but device complexity and resource requirements deteriorate
Solution Approach 1:
The patent creates a universal AI-based system that can handle multiple optimization problems across different domains using a single large language model, reducing the need for multiple specialized expert systems and simplifying overall system complexity while maintaining high accuracy
Data Source
AI summary
An approach for generating optimization solutions may be presented herein. The approach may include generating an optimization solution verification program. The optimization solution verification program code can be automatically converted into a loss function, where the objection function constraints associated with the optimization solution verification program are incorporated into the loss function. A plurality of random inputs for the optimization issue can be generated and used to train a sequence generation model, based on the generated loss function. An optimized solution can be generated with the trained sequence generation model.