FICO Xpress Optimization Suite

• FICO Xpress is a comprehensive optimization suite supporting LP, MIP, NLP, and global optimization with state-of-the-art algorithms.
• It implements advanced techniques like branch-and-bound with cutting planes, parallel simplex methods, and spatial branch-and-bound for efficient solving.
• The suite features the Mosel modeling language and machine learning-aided solver decisions to enhance configuration and performance.

FICO Xpress is a commercial suite for mathematical optimization encompassing mixed-integer programming (MIP), linear programming (LP), quadratic and general nonlinear programming (NLP), and global optimization. The suite comprises advanced solvers, a modeling language (Mosel), and developer APIs. It is widely deployed in industrial, scientific, and academic applications, due to both its solution performance and extensible modeling environment.

1. Core Algorithms and Solver Architecture

FICO Xpress implements state-of-the-art algorithms for linear, mixed-integer, and nonlinear optimization, integrating several proprietary and open research advances. For LP and MIP:

• Branch-and-Bound with Cutting Planes: The branch-and-bound (B&B) framework forms the core of the MIP solver. At each node $t$ of the B&B tree $\mathcal{T}$, Xpress solves an LP relaxation $Q_t$, temporarily relaxing integrality to obtain fractional solutions.
• Cutting Plane Management: Xpress supports global and local separation of valid inequalities (cuts) of the form $\alpha^\top x \leq \beta$ to tighten the LP relaxation. Global cuts are added at the root node, while local cuts can be generated at subtrees to strengthen bounds locally. The integration, scheduling, and cost-benefit analysis of such cuts is central to solver performance (Berthold et al., 2022).
• Parallel Simplex Engines: For general large-scale sparse LPs, Xpress incorporates two parallel dual revised simplex methods, PAMI (Parallel Across Multiple Iterations) and SIP (Single Iteration Parallelism). PAMI uses suboptimization-based pivoting with candidate sets for major/minor iterations, allowing for parallel FTRAN (forward transformation) tasks via the APF (Alternate Product Form) update. SIP overlaps the computational phases of each simplex iteration, exploiting per-iteration concurrency (Huangfu et al., 2015).

For nonlinear and global optimization:

• Spatial Branch-and-Bound: Xpress Global runs a spatial branch-and-bound algorithm on continuous variables. Factorable nonlinear structures are automatically detected and relaxed using McCormick envelopes (for bilinear and quadratic terms) and smooth convex function approximations.
• Outer-Approximation and Local Search: For nonconvex models, the solver employs outer-approximation (OA) on nonlinear equalities and primal heuristics such as local NLP solves and feasibility pumps for MINLP contexts (Berthold et al., 9 Jan 2026).

2. Machine Learning-Aided Solver Decisions

Recent versions of Xpress integrate learned decision-support mechanisms for solver configuration. For example, Berthold, Francobaldi, and Hendel (2022) introduced a regression forest model trained to predict, based on 32 features comprising both problem structure and dynamic root-node information, whether local cuts would speed up or slow down the solution for given problem instances.

The learned model predicts the log-speed-up $$y=\log_2\left(\frac{\text{Time}_{LC}+1}{\text{Time}_{NLC}+1}\right)$$ and triggers local cut deactivation when the predicted benefit is negative. This is realized by embedding the model into Xpress's C API and cut-management callback system. On industrial instance sets, the learned approach achieves up to 5.3% total speed-up on hard instances, and up to 18% on affected models, with statistical significance (Wilcoxon signed-rank $p<0.001$) (Berthold et al., 2022).

3. Modeling Environment: Mosel

Xpress Mosel is the proprietary modeling language supporting algebraic definition of MILP, MINLP, and general nonlinear models. Key features:

• Symbolic Declaration: Support for sets, parameters, variables, and constraints declaration. Mosel offers fine-grained control over data-driven model instantiation, array-based variables, and dynamic structures.
• Integration with Data Sources: Direct integration with ODBC and Excel for data import/export, facilitating industrial-scale workflows. SQL queries and Mosel helper functions abstract input/output for large and dynamic datasets.
• Dynamic Model Construction: Features such as logic for optional variables, custom helper functions (e.g., for startup cost curve thinning in power market models), and assertion-based error trapping support robust implementations (Brandenberg et al., 2014).

4. Performance and Empirical Results

Benchmarks on industrial and synthetic test sets demonstrate that FICO Xpress offers robust performance across a spectrum of problem types:

• Parallel Dual Simplex: Average speedups of 1.5–2× on large LPs (relative to serial solvers) using 8-core commodity CPUs, with PAMI and SIP engines auto-selected depending on matrix sparsity (Huangfu et al., 2015).
• MIP Solver (Local Cuts): For MIPLIB 2017, machine-learned local cut control achieved a 3.3% time speed-up on unseen instances; for hard cases (solve time ≥ 1000 s), speed-up reached 18.7% (Berthold et al., 2022).
• Global Optimization: In nonconvex NLPs drawn from combinatorial geometry, default settings without custom intervention produced solutions matching or exceeding best-known results, with run times in seconds to minutes for models up to 91 variables and 876 constraints (Berthold et al., 9 Jan 2026).
• Power Market Models: In real-world unit commitment implementations with 50 generators and 168 time periods, solve times were typically 10–30 s (4-core CPU), and cost curve thinning yielded further solve-time reductions (Brandenberg et al., 2014).

5. Industrial Use Cases and Applications

FICO Xpress is deployed in diverse optimization scenarios:

• Power Market Modeling: Fully featured unit commitment models in Mosel, supporting binary on/off, ramping, storage, technical constraints, and multi-source data integration (Brandenberg et al., 2014).
• Combinatorial Geometry: Direct solution of nonconvex packing and distance-ratio problems, as demonstrated on AlphaEvolve benchmarks for geometric arrangements (Berthold et al., 9 Jan 2026).
• Large-Scale Mixed-Integer Programs: Broad applicability in logistics, scheduling, finance, and supply chain, benefiting from the automated cut generation, presolve routines, and user-integration via APIs.

6. Solver Configuration, Tuning, and Extensibility

The suite exposes detailed solver parameters for tuning:

• Options: Control over threading (XPRS_threads), time limits, MIP focus, tolerances, and presolving. These provide users with extensive controls for balancing optimality, precision, and computational resources.
• Automatic Heuristics: Autoselection of parallel simplex engine (PAMI or SIP), learned cut management, and primal heuristics tuned to model class and data-driven characteristics.
• Low-Level Integration: The C API, advanced callbacks (node, cut, branch), and user-defined heuristic support permit extension and embedding of custom logic, or integration into large decision-support systems.

7. Research Impact and Future Directions

FICO Xpress serves as a reference point for optimization research due to:

• Translating Research into Practice: Fast adoption of community innovations (suboptimization, parallel LP, automatic cut management) positions Xpress as a benchmark for both industrial and comparative academic studies.
• Extensible Decision Models: Current ML-based policies focus on binary decisions (local cuts everywhere vs. at root only); planned extensions target adaptive, depth-dependent decisions, cut family selection, and conflict analysis integration (Berthold et al., 2022).
• Robust Solution Technology: Xpress matches or outperforms both heuristic and LLM-discovered solutions for highly nonlinear, constrained optimization problems, with minimal specialized tuning (Berthold et al., 9 Jan 2026).

In summary, FICO Xpress is a comprehensive, multifaceted optimization platform distinguished by high-performance parallel algorithms, advanced cut and node management, tight integration with modeling and data-handling environments, and demonstrated impact on both academic and industrial optimization practice.