When discussing optimization solution development, I often refer to “the journey” from opportunity identification and business problem framing to deployment and solution lifecycle management, and the important steps in between. As a consulting firm, we are engaged by organizations whose in-house teams can be at varying stages of the journey with a given solution.
This diagram maps representative projects to where we successfully partnered with our clients. It is based on the INFORMS Analytics Framework, formerly called the Job Task Analysis (JTA), https://www.certifiedanalytics.org/jta. I was one of a small group of experts that revised the JTA in 2024, updating the original version created in 2012. Its seven stages are depicted in the large central arrow. For success with any significant optimization solution, all seven stages must be adequately performed. In other words, it is inadvisable to focus heavily on model building without accounting enough for the other stages—a common occurrence.
Business Problem Framing
At the beginning of the framework is the determination of a business question whose answer brings value to the organization. For Princeton, this initial activity can take place in a series of workshops, field observations and meetings to establish a problem statement and usability requirements, and to identify the stakeholders. Is the problem amenable to an analytics solution? If it is, the parties refine the problem statement and delineate constraints. Working with business and technical executives, the potential solution’s benefits and value are defined. The value is often savings from greater efficiency, but it may also be improved yield or capacity, or other metrics such as a reduction in the time to make critical decisions. Obtaining agreement from the stakeholders at the end of this stage is foundational to later success. For more on this stage, see this previous post and this post.
Analytics Problem Framing
The business problem is reformulated into an analytics problem with a potential analytics solution. The team develops a proposed set of drivers and relationships to outputs and states the related set of assumptions. Key metrics of success are defined, with the requirement that they be measurable. For tips about KPIs for optimization solutions and obtaining stakeholder agreement on the approach, see this post by my colleague Steve Sashihara for tips.
As the diagram above shows, we at Princeton have been engaged to assist during or after the initial two stages have been conducted.
Data
For several years in live workshops at INFORMS meetings and elsewhere and in webinars, I have encouraged a “Data First” approach rather than a “Model First” approach in optimization solution development. See this recent post on the topic. I cannot stress enough the importance of identifying and prioritizing data needs and sources; acquiring, harmonizing, rescaling, cleaning, and sharing data; and identifying relationships in the data. Documentation is critical. The problem statements may be refined at this stage. For more, see this post by my colleague Rob Randall.
Methodology (Approach) Selection
This stage entails the identification of potential approaches for solving the business problem, and the selection of software tools. In our engagements, we have varying degrees of freedom with the toolkit: the client may require a language like Python or a specific solver, which we observe. The current optimization software landscape includes a variety of tools besides the solver which can be helpful and should be considered, for example, with UI development. See this post for an example. Experimenting and testing of different approaches can be done with prototypes.
Model Building
Traditionally, modeling is the bread and butter of mathematical optimization education, although the INFORMS Analytics Framework makes it clear that so much more is required for solution success. Identifying suitable model structures, running and evaluating the model, calibrating the models and data, and integrating the models are all critical. Documentation of the mathematical model is important for the long term: we at Princeton have been engaged numerous times to recreate, translate and/or upgrade a model working mainly or exclusively off the documentation because the code is not available or the original developers had moved on or retired.
Assistance in development and testing may be required for a model that has yet to be deployed. There are also occasions when an existing model calls for upgrading due to performance issues or changes in the business requirements. Princeton offers a model review service, described in this post, which has led to many good outcomes. Sometimes an expert third party is essential for such quality assurance and improvement.
Deployment
At Princeton we typically equate engagement success with the deployment of the optimization model into production. Employing an iterative “Agile for Optimization” development methodology is one way to improve the likelihood of implementation, as my colleague Patricia Randall recently wrote. Steve Sashihara has given many presentations on overcoming challenges to full field installations, as described in this post.
Deployment involves two important tasks. One part is the technical aspect of creating a software application for use by our clients, whether in an on-prem environment or in the cloud. The second part, often overlooked by many optimization practitioners, is the deployment of the solution into the operations of the business. How will the people at our clients be affected by this solution? Are there downstream effects in business operations? Large-scale deployments merit consideration of Change Management strategies, which I have written about.
Model Life Cycle Management
Managing the model life cycle and perpetuating its business benefit over time typically requires a team. Most modelers do not want to stay involved; they want to solve new problems! Training and documentation are required. Beyond maintenance, at Princeton we support optimization models that call for continual incremental improvements that yield corresponding incremental increases in value because of the nature of the business process.
Conclusion
As demonstrated in the diagram above, Princeton has been involved in numerous projects covering the various aspects of the INFORMS Analytics Framework. Our best practices have been developed over decades and are also described in Steve Sashihara’s book, The Optimization Edge. These best practices have now been codified in the INFORMS Analytics Framework, on which I had the pleasure and honor of working with a great set of INFORMS colleagues to update in 2024. If your projects are not considering all the domains of the framework, they are likely to struggle to succeed. Contact us to find out how we can help you along the journey with your optimization projects to maximize their business value.
