Introduction
Generally, optimization practitioners believe the most important feature for solver selection is performance. In practice, most Data/Decision Science (DS) teams measure this by time to optimality, gap within a time limit, and finding high-quality solutions early in the search.While solve performance is indeed important when solving a particular optimization problem, it remains myopic from an enterprise perspective. Companies don’t create business value based on the speed with which each solve runs, rather, how fast its team can create decision assets that lead to desired outcomes. In the bigger picture, factors such as capabilities, time to value, total cost of ownership, user control, and reproducibility, play a larger role in a company’s success.
In this blog post, we’ll present recommendations on how to select the solver that not only solves the current optimization model fast but also enables your DS team to generate the most business value for your organization.
TLDR: The table below summarizes our recommendations and suggested criteria.
| Characteristic | Recommendations/Criteria |
| Capabilities |
|
| Solver Performance |
|
| Time to Value |
|
Evaluating a Solver
Capabilities
One of the most important capabilities to consider when selecting a solver is what optimization problem types it’s able to solve. To classify the types of optimization models your company solves, you can leverage the Neos webpage’s taxonomy of algebraically expressible optimization problems. FICO Xpress can solve a wide breadth of optimization models to either global or local optimality. These include Mixed Integer, Quadratic, Mixed Integer Quadratically Constrained, Second Order-Cone, Non-linear, and Mixed-Integer Non-linear Programs as well as Blackbox optimization and Constraint Programming problems. This broad scope allows teams to freely build models without compromising due to solver capabilities.The second most important capability to consider is the languages through which you can interact with the solver. Users can interact with FICO Xpress through Python, .NET, C/C++, Java, MATLAB and R as well as through algebraic modelling languages such as AMPL, GAMS, and Xpress Mosel, making it easy to integrate into almost any workflow.
User control is another important capability that you should consider in your evaluation. For tailored algorithms meant to exploit every bit of performance from known problem structure, the level of control that users have over the solve process is critical. FICO Xpress offers a comprehensive set of callbacks for advanced user control, even allowing users to influence how subproblems are created during branching.
Figure 1: Summary of main MIP related callbacks in FICO Xpress
Finally, reproducibility is another key capability, especially in cases where the development and production environments are different or consistently change as in ephemeral environments like containers. With determinism, reproducing solve paths across different environments is simplified, allowing more streamlined development and troubleshooting and more stable performance across different compute resources. FICO Xpress is currently the only solver in the market that guarantees determinism not only across operating systems, but also across hardware. If your company uses ephemeral environments for development and deployment, this should be an important characteristic to consider.
Solver Performance
Despite what solver marketing materials suggest, no single optimization solver can be the fastest on all problems. In fact, even the same solver’s performance can vary over different types of models. The practical reality is that optimization models will have different underlying mathematical structure requiring different trade-offs between exploration versus exploitation. The same solver-routines that work well under specific mathematical structures might cause performance degradation for others.With FICO Xpress, users can tailor the solver’s search strategy through control parameters. This process, known as solver tuning, has been shown to make solves up to 60x faster in some instances. FICO Xpress users can easily tune the solver themselves on either a single instance or a set of model instances using the automated tuner which offers a range of target performance metrics to tune. In addition, FICO Xpress offers a FREE tuning service performed by customer-focused, friendly experts with decades of experience and advanced degrees.
Figure 2: Examples of solver tuning success stories
Executing multiple solvers
Though evaluating solve speed requires investing time running each solver, it does not require having to rewrite your model for each solver API. Thanks to industry-standard output formats like MPS and LP matrix files, one can write their models in their preferred API, export it to any of these industry standard outputs, and evaluate it with any optimization solver. In FICO Xpress, the writeProb, readProb, and tune functions enable seamless performance evaluation across different solvers by exporting models to these universal formats through their favourite API language or the optimizer console.Figure 3: Using the optimizer console to tune an instance
Alternatively, practitioners can use modeling frameworks (Pyomo, JuMP, MATLAB) and algebraic modeling languages (AMPL, GAMS) that provide unified APIs with configurable solver backends. These are particularly helpful for evaluations that require iterative executions to solve a model, granted that the necessary API calls exist. However, once the evaluation process is complete, using the solver’s native APIs can provide a more seamless experience for production-grade development.
Finally, all solves related to the evaluation should be run on the same or similar servers/computers capable of handling the size, type, and required speed with which you’ll run your optimization model in production to avoid skewing the results due to different hardware performance.
Time to Value
An important factor that is often overlooked during solver selection is Time to Value which measures how long it takes for technology projects to provide business value, offsetting their cost and bringing returns on the initial investment. Common metrics are time from development to deployment, and total annualized cost of infrastructure, licenses, development, and maintenance, i.e. Total Cost of Ownership.Beyond just a state-of-the-art solver, FICO Xpress also developed a free algebraic programming language, Xpress Mosel, and a deployment platform, Xpress Insight. Mosel allows writing models using language close to the algebraic representation but with the performance of efficient C-code. This makes writing optimization models more intuitive for those with less software development experience, accelerating development speed and facilitating model maintenance.
FICO Xpress’s deployment platform, Xpress Insight, allows DS teams to build and deploy scalable enterprise applications using Mosel or Python. It removes the dependence on software engineering and IT resources to deploy rich decision support applications and has been shown to reduce development and deployment time to a matter of days instead of months. It simplifies the web application development process to adding XML configurations and markup to your optimization models without the need for HTML/ Javascript knowledge.
Figure 4: Example of Insight deployed application
With Xpress Insight, IT teams are brought in only for the initial setup after which DS teams can build, deploy, and maintain multiple optimization application services to as many stakeholders/ departments as requested. In addition to reducing development time, this setup significantly decreases the total cost of ownership of optimization-based decision assets for your organization, compounding value with every new decision asset that is deployed.
Conclusion
Choosing the right optimization solver goes far beyond finding the fastest option for individual problems. It requires a strategic approach that matches technology with long-term business goals. The most successful companies understand that solver selection is a foundational decision that impacts every future optimization project, affecting not just speed but the entire growth of their Data/Decision Science capabilities. By focusing on comprehensive features, strategic time-to-value benefits, and overall cost of ownership alongside traditional performance measures, organizations such as PepsiCo, Southwest Airlines, DoorDash, and Procter & Gamble have set themselves up for significant returns on their optimization investments. Try it yourself for freeThe real measure of success isn't in milliseconds saved per solve, but in the company's improved ability to adapt, scale, and deliver clear business results through better decision-making tools that create lasting competitive advantages.
Updated 9 months ago
Version 3.0