Accelerated Supply Chain Optimization and Decision-Support in Contested Environments
Abstract
This proposal builds on past work developing FullSolve, a mixed integer linear programming model that computes detailed logistics support plans for military operations within a larger effort to provide logistics decision support. This larger effort is called Visual Integrated Tactical Logistics # Battle Management Aid (VITL-BMA). While FullSolve has successfully delivered value to stakeholders, greater value would be provided by achieving three key goals: reducing solve times to below a threshold 20 minutes, making the product easier to use by non-experts, and expanding the number of real-world phenomena that it models.To achieve these goals, we propose three parallel research thrusts:First, we will accelerate FullSolve s core optimization capabilities through multiple complementary approaches: (1) Code optimization targeting bottlenecks in data processing while constructing formulations and when solving mixedinteger programs, (2) Automated configuration of generalized versions of two effective heuristics, hub-spoke and sea-air split, using graph neural networks, and (3) Dual decomposition methods that partition the problem into independently solvable components. Early results show our decomposition approach matches optimal solutions on test problems while promising better scaling to larger instances.Second, we will develop an AI-powered interactive decision support system that makes FullSolve accessible to military planners who are experts in logistics but not optimization. The system will automatically translate operational goals into optimization parameters through natural interaction, pre-compute diverse solution sets using Bayesian optimization, and support the presentation of results through an intuitive faceted search interface. By incorporating user feedback and interaction patterns, the system will continuously improve its ability to anticipate user needs and generate relevant solutions.Third, we will expand FullSolve s modeling capabilities through a novel "guess-and-check" framework that enables integration with specialized models developed by others working on VITL-BMA. Rather than directly implementing complex phenomena like crew rest requirements or adversarial disruption - which would substantially increase computational complexity - FullSolve will generate solutions using simplified constraints. External validation services will then verify if solutions satisfy detailed requirements, with violations translated into updated constraints in an iterative process. This microservice-style architecture enables domain experts to contribute independently while maintaining FullSolve s computational efficiency.The proposed work will enhance FullSolve#s capabilities to create an accessible platform that can support rapid, interactive logistics planning while leveraging the full capabilities of the VITL-BMA team. Success will be measured through improved solve times, increased adoption by non-expert users, and expanded modeling capabilities validated through tabletop exercises.Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 12, 2025
- Source ID
- N000142512181
Entities
People
- Peter Frazier
Organizations
- Cornell University
- Office of Naval Research
- United States Navy