931: Boost Your Profits with Mathematical Optimization, feat. Jerry Yurchisin

Gurobi's Jerry Yurchisin explains the power of mathematical optimization, a prescriptive approach that complements AI's predictive capabilities. This summary covers how to get started with free resources, the use of GPUs and LLMs to enhance optimization, real-world applications at companies like Toyota, and its relationship with quantum computing.

October 14, 2025 5 min read

931: Boost Your Profits with Mathematical Optimization, feat. Jerry Yurchisin — source: https://www.youtube.com/watch?v=5yxg1pN1yh8

The Prescriptive Power of Mathematical Optimization

While traditional machine learning and AI excel at prediction—forecasting what the future might look like—they don’t prescribe the exact actions to take in response to those predictions. Mathematical optimization fills this critical gap. It provides a framework to find the best possible outcome by modeling a business problem through three core components:

Decision Variables: The levers you can pull or choices you can make (e.g., which facilities to open, how much product to ship).
Constraints: The business rules, physical limitations, or regulations that restrict your decisions (e.g., budget limits, production capacity).
Objective Function: The goal you are trying to achieve, which you want to either maximize (like profit) or minimize (like cost or carbon footprint).

The output is not a prediction, but a concrete set of optimal decisions. For instance, it can determine “how much of product A should I make at this facility and ship to that distribution center?” The complexity of modern business, with thousands of products, suppliers, and customers, creates a decision space that is too vast for human intuition or simple predictive models to navigate effectively.

Bridging the Gap: GPUs and Large-Scale Problems

Historically, the algorithms used in mathematical optimization, particularly the simplex algorithm, involved an operation called “pivoting,” which is not well-suited for the highly parallel architecture of GPUs. This made CPUs the traditional hardware of choice.

However, Gurobi, in partnership with NVIDIA, is exploring new frontiers to leverage GPU power. For certain problem types, specifically very large-scale Linear Programs (LPs), newer algorithms can take advantage of GPU acceleration to deliver significant reductions in solve times. This collaboration was highlighted by NVIDIA’s CEO Jensen Huang at GTC and is exemplified by NVIDIA’s open-source co-opt solver. Gurobi is now integrating GPU support into its own products, enabling users to tackle previously intractable, massive-scale optimization problems.

Getting Started: Free Learning Resources

Gurobi has made a concerted effort to make mathematical optimization accessible to the data science and AI community. A wealth of free resources is available at gurobi.com/learn. Key starting points include:

Free License: By simply running pip install GurobiPy, users get access to a free, size-limited license perfect for learning and prototyping.
Notebooks and Tutorials: Jerry Yurchisin has personally created numerous free tutorials and code notebooks that guide users through practical examples.
University Partnerships: A comprehensive four-part Udemy course, “Introduction to Optimization Through the Lens of Data Science,” was developed in collaboration with Georgia Tech.
Video Series: A multi-part YouTube series (OPI 101, 201, 202) provides in-depth training, featuring not just Gurobi experts but also industry consultants who share real-world modeling perspectives.

These resources are designed for data scientists who may not have a traditional Operations Research background, demonstrating that the field is accessible to a broader technical audience.

The Guarantee of Optimality

A key differentiator of mathematical optimization is its ability to deliver a provably optimal solution. Unlike heuristics or many machine learning models that find “good” solutions, optimization solvers use rigorous mathematical proofs to guarantee that no better solution exists, given the model’s constraints and objective.

Even for extremely complex problems where finding the proven optimum might take too long, the solver provides a valuable metric called the MIP gap. This gap quantifies the worst-case difference between the current best solution and the theoretical optimum. Knowing you are within, for example, 2% of the best possible outcome is a powerful piece of information that no other approach can provide. This is possible due to the underlying geometry of the problem space, which allows the algorithms to systematically explore and eliminate entire regions of suboptimal solutions.

The Modern Frontier: LLMs and AI in Optimization

To lower the barrier to entry and speed up the development process, Gurobi is integrating Large Language Models (LLMs) into the optimization workflow. They have released several tools in the ChatGPT store, including:

AI Modeling Prompt Engineer: A conversational tool that helps users translate a vague business problem into a well-defined problem statement, clarifying the objective, decisions, and constraints.
AI Modeling Assistant: Takes the structured problem statement and generates a first draft of both the mathematical formulation and the corresponding Python code using the GurobiPy library.

Gurobi is also developing its own self-hosted tool, “Groot,” to further assist practitioners. These tools act as co-pilots, helping users—especially those new to the field—to correctly formulate their problems and accelerate the coding process, making the power of optimization more accessible than ever.

Real-World Applications and Use Cases

Toyota (Vehicle Manufacturing): Toyota developed a sophisticated planning tool where an optimization model forms the core. Planners, who are not optimization experts, can interact with the model through a natural language (LLM) interface. They can run “what-if” scenarios to understand the impact of events like fluctuating tariffs or supply chain disruptions, allowing them to make robust, data-driven manufacturing decisions.
Total Wine (Inventory Management): This case highlights the journey of a data science team with no prior operations research experience. They successfully built and deployed a mathematical optimization model to manage their complex inventory. The problem involved countless decisions, from which brands of beer and wine to stock, to the specific quantities of different package sizes (e.g., six-packs vs. cases). The project was a massive success, leading to significant cost savings that could be reinvested into the business.

Future-Proofing: Mathematical Optimization vs. Quantum Computing

There’s a common misconception that today’s hardest problems are unsolvable and must wait for the arrival of quantum computing. However, this overlooks the immense power of current mathematical optimization solvers. The argument is not to wait, but to act now:

Immediate Value: Many problems that people earmark for quantum computing can be solved today with mathematical optimization, delivering immediate ROI.
Essential Foundation: The most challenging part of any optimization project is the foundational work: deeply understanding the business problem, defining constraints, gathering data, and securing stakeholder buy-in. This work must be done regardless of whether the back-end solver is classical or quantum.
Future-Ready: By implementing mathematical optimization now, organizations build the necessary expertise and establish a reliable performance benchmark. When quantum computing becomes commercially viable, they will be perfectly positioned to adopt it, comparing its performance against a proven, optimal baseline rather than starting from scratch.