Current Position
Assistant Professor of MS&E and, by courtesy, of Electrical Engineering
Intersection with the project
Overview
Professor Madeleine Udell’s research offers an essential technical and sociotechnical bridge for the agricultural management chatbot project, especially regarding the democratization of access to complex optimization tools. Her work addresses both the development of artificial intelligence systems capable of translating natural language into rigorous mathematical models and the understanding of the barriers to adoption of these tools by non-specialist users. In the context of the project, Udell’s work validates the choice of a conversational interface (WhatsApp) to handle real-world data (“messy data”) and transform the rural producer’s informal routine into optimized and reliable management decisions.
”OptiMUS: Scalable Optimization Modeling with (MI)LP Solvers and Large Language Models” (2024)
This research introduces OptiMUS, an agent based on Large Language Models (LLMs) that automates the formulation and solving of optimization problems from natural language descriptions. The system uses an agent-based framework (Manager, Formulator, Programmer, Evaluator) to overcome limitations of common AI models, such as processing long contexts and complex data. The term (MI)LP refers to (Mixed Integer) Linear Programming, a mathematical technique used to find the best solution among several alternatives, such as optimizing routes or the use of inputs
Translating Natural Language into Action
- “OptiMUS, a Large Language Model (LLM)-based agent designed to formulate and solve (mixed integer) linear programming problems from their natural language descriptions.”. “OptiMUS, a Large Language Model (LLM)-based agent designed to formulate and solve (mixed integer) linear programming problems from their natural language descriptions.”
- Intersection: The chatbot uses Advanced Audio Processing to perform semantic interpretation of informal audios sent by the producer. Udell’s work provides the technical basis for these voice commands to be transformed by the AI into triggers and mathematical models for operational execution
Accessibility for Small Businesses and Sectors without Specialists
- “Automating optimization modeling would allow sectors that cannot afford to have access to optimization experts to improve efficiency using optimization techniques.” “Automating optimization modeling would allow sectors that cannot afford to have access to optimization experts to improve efficiency using optimization techniques.”
- Intersection: The chatbot’s target audience is rural producers who prefer “practical operations” and do not have operations research specialists. The system acts as the automated “executive secretary” that brings the efficiency of optimization into the farm without the need for advanced technical knowledge on the user’s part
”“It Was a Magical Box”: Understanding Practitioner Workflows and Needs in Optimization”
This is a qualitative study that analyzes the real workflows of optimization model developers (OMDs). Udell and her collaborators propose the “Three Ds” (Data, Decisions, Dialogue) as the pillars for the success of optimization projects in the real world. The term “Dialogue” refers to continuous communication with stakeholders to build trust, while “Problem Elicitation” describes the phase of translating vague business needs into technical requirements.
Importance of Dialogue for Trust and Adoption
- “Our findings reveal that optimization practice is not only about algorithms that deliver better decisions, but is equally shaped by data and dialogue—the ongoing communication with stakeholders that enables problem framing, trust, and adoption.” “Our findings reveal that optimization practice is not only about algorithms that deliver better decisions, but is equally shaped by data and dialogue—the ongoing communication with stakeholders that enables problem framing, trust, and adoption.”
- Intersection: The project chooses WhatsApp as the main interface due to its “high familiarity.” Udell demonstrates that optimization fails if it is an opaque “magic box”; by using a daily dialogue platform, the chatbot builds the trust needed for the rural producer to accept the AI’s alerts and suggestions
Handling “Messy” Real-World Data
- “optimization is characterized by messy and incomplete data that inform and constraint model formulation…” “optimization is characterized by messy and incomplete data that inform and constraint model formulation…”
- Intersection: The project’s Continuous Learning Engine handles data extracted from audios, messages, and receipt images, which are inherently informal and “messy.” Udell highlights that 70% of the effort in real projects is in processing this data, validating the project’s technological focus on transforming informal field routines into structured data for management
Full papers
Madeleine Udell - OptiMUS Scalable Optimization Modeling with (MI)LP Solvers and Large Language Models.pdf Madeleine Udell - “It Was a Magical Box” Understanding Practitioner Workflows and Needs in Optimization.pdf