AI Prescriptive Policy Generation with Interdependent Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current prescriptive analytics struggle to generate accurate and constraint-satisfying decision paths due to the complexity of interdependent operational information and large numbers of possible decision paths, often resulting in poor performance and rule conflicts.
Innovation Solution
A computer-implemented method using AI-driven prescriptive policies is developed, where a model is trained with interdependent operational information to dynamically generate decision paths and apply mixed-integer programming for constraint satisfaction, incorporating a teacher and student model with multiway-split trees to identify optimal policies that satisfy attribute-combination and global capacity constraints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional prescriptive analytics methods are used to generate decision paths, then the system can handle simple constraints, but it fails to accurately satisfy complex interdependent operational constraints and produces rule conflicts
Solution Approach 1:
The patent segments the complex constraint satisfaction problem into two distinct phases: (1) generating candidate decision paths using a teacher model that predicts counterfactual outcomes, and (2) selecting optimal paths using a student model with mixed-integer programming that explicitly handles interdependent constraints. This segmentation allows each phase to specialize, improving overall reliability without overwhelming system complexity.
Solution Approach 2:
The patent introduces an intermediary mixed-integer programming formulation that acts as a bridge between the teacher model's candidate paths and the final prescriptive policy. This intermediary layer with dynamic column generation systematically evaluates candidates against interdependent constraints, resolving rule conflicts while maintaining manageable complexity through mathematical optimization.
2Measurement precision
If the system evaluates all feasible decision paths to ensure constraint satisfaction, then decision accuracy improves, but processing power and storage requirements increase significantly
Solution Approach 1:
The teacher model performs preliminary action by pre-generating and filtering candidate decision paths before the optimization phase. This preliminary filtering reduces the search space from all feasible paths to a manageable subset of promising candidates, thereby maintaining decision accuracy while significantly reducing the processing power needed for subsequent constraint evaluation.
Solution Approach 2:
Instead of evaluating all feasible decision paths, the system performs partial evaluation by focusing computational resources on the most promising candidate paths identified by the teacher model. The mixed-integer programming formulation with dynamic column generation evaluates only necessary candidates, avoiding excessive computation while maintaining precision through targeted analysis of high-probability paths.
3Measurement precision
If the model generates highly accurate prescriptive policies, then the policies become less interpretable, but interpretability is crucial for user understanding and trust
Solution Approach 1:
The student model with mixed-integer programming serves as an intermediary that translates the teacher model's accurate but opaque predictions into an interpretable prescriptive policy. By formulating the selection process as a structured optimization problem with explicit constraints and objective functions, the system maintains high accuracy while producing transparent, explainable decision paths that users can understand and trust.
Solution Approach 2:
The patent segments the modeling process into a black-box teacher model for accurate prediction and a white-box student model for interpretable policy generation. This segmentation allows the system to leverage the strengths of both approaches: the teacher model provides accurate counterfactual predictions, while the student model translates these into interpretable prescriptive policies with clear logical structure and explicit constraint satisfaction.
Data Source
AI summary
A computer-implemented method of generating an Artificial Intelligence (AI) driven prescriptive policy and executing a function includes obtaining interdependent operational information about the function. A model is trained with the interdependent operational information about the function to dynamically generate a plurality of candidate decision paths from a group of all feasible decision paths for a plurality of interrule logical conditions and one or more dynamic constraints of the operational information. A prescriptive policy is generated from the plurality of candidate decision paths to execute the function that satisfies to a threshold degree of confidence the interrule logical conditions and the one or more dynamic constraints of the operational information. The function is executed based on the generated prescriptive policy.