HiGHS - High Performance Optimization Software

Comprehensive assistance with HiGHS development, generated from official documentation.

HiGHS is open-source software for defining, modifying, and solving large-scale sparse linear optimization models. It's freely available under MIT license with no third-party dependencies.

When to Use This Skill

Use this skill when:

• Solving linear programming (LP) problems
• Solving mixed integer linear programming (MILP) problems
• Solving quadratic programming (QP) problems
• Implementing optimization solutions in C++, Python (highspy), Julia, C, C#, Fortran, or Rust
• Building or modifying optimization models programmatically
• Working with MPS or CPLEX LP file formats
• Optimizing performance with parallel solvers or GPU acceleration
• Setting solver options and tolerances
• Hot-starting from existing solutions or bases
• Extracting model information and solution data
• Modifying existing optimization models

Problem Types Supported

HiGHS can solve problems of the form:

Linear Programming (LP):

minimize    c^T x
subject to  L ≤ Ax ≤ U
            l ≤ x ≤ u

Mixed Integer Linear Programming (MILP): Same as LP, but some variables must take integer values.

Quadratic Programming (QP): LP with additional objective term ½x^T Q x where Q is positive semi-definite. (Note: Cannot solve integer QP problems)

Solvers Available

• Simplex methods: Revised simplex (primal and dual) - Most robust for general LP
• Interior point method: Two implementations (IPX serial, HiPO parallel)
• PDLP: First-order primal-dual method (GPU-accelerated option available)
• Branch-and-cut: For MILP
• Active set: For QP

Quick Reference

1. Basic Setup and Solve (Python)

Initialize HiGHS and solve a model from file:

import highspy
import numpy as np
h = highspy.Highs()

# Read a model from MPS file
filename = 'model.mps'
status = h.readModel(filename)
print('Reading model file', filename, 'returns a status of', status)

# Solve the model
h.run()

# Get solution
solution = h.getSolution()
info = h.getInfo()

2. Building a Simple Model (Python Simplified Interface)

Build an optimization model programmatically:

# Problem:
#   minimize    f  =  x0 +  x1
#   subject to              x1 <= 7
#               5 <=  x0 + 2x1 <= 15
#               6 <= 3x0 + 2x1
#               0 <= x0 <= 4; 1 <= x1

import highspy
h = highspy.Highs()

x0 = h.addVariable(lb = 0, ub = 4)
x1 = h.addVariable(lb = 1, ub = 7)

h.addConstr(5 <=   x0 + 2*x1 <= 15)
h.addConstr(6 <= 3*x0 + 2*x1)

h.minimize(x0 + x1)

3. Building a MILP Model (Julia C API)

Complete MILP example using Julia's C API wrapper:

using HiGHS

highs = Highs_create()
ret = Highs_setBoolOptionValue(highs, "log_to_console", false)
@assert ret == 0  # If ret != 0, something went wrong

# Add columns (variables)
Highs_addCol(highs, 1.0, 0.0, 4.0, 0, C_NULL, C_NULL)   # x is column 0
Highs_addCol(highs, 1.0, 1.0, Inf, 0, C_NULL, C_NULL)   # y is column 1

# Set y as integer variable
Highs_changeColIntegrality(highs, 1, kHighsVarTypeInteger)

# Set objective to minimize
Highs_changeObjectiveSense(highs, kHighsObjSenseMinimize)

# Solve
Highs_run(highs)

4. Using with Julia JuMP (High-Level Interface)

using JuMP
import HiGHS

model = Model(HiGHS.Optimizer)
set_optimizer_attribute(model, "presolve", "on")
set_optimizer_attribute(model, "time_limit", 60.0)

# Define your optimization model...
@variable(model, x >= 0)
@variable(model, y >= 0)
@objective(model, Min, x + y)
@constraint(model, 5 <= x + 2*y <= 15)

optimize!(model)

5. Extracting Solution Values Efficiently (Python)

Important: Direct array access is slow in Python. Convert to list first!

import highspy
h = highspy.Highs()
h.readModel('model.mps')
h.run()

# Get solution object
solution = h.getSolution()

# SLOW: Accessing directly from solution.col_value
# for i in range(num_cols):
#     val = solution.col_value[i]  # Takes 0.04s

# FAST: Convert to list first
col_value = list(solution.col_value)
for i in range(num_cols):
    val = col_value[i]  # Takes 0.0001s (400x faster!)

6. Setting Options and Choosing Solver

import highspy
h = highspy.Highs()

# Set common options
h.setOptionValue("presolve", "on")
h.setOptionValue("time_limit", 100.0)
h.setOptionValue("mip_rel_gap", 0.01)

# Choose specific solver
h.setOptionValue("solver", "simplex")  # or "ipm", "pdlp", "hipo"

# For GPU acceleration with PDLP
h.setOptionValue("solver", "pdlp")
h.setOptionValue("kkt_tolerance", 1e-4)  # Recommended for PDLP

7. Command Line Usage (Executable)

# Basic solve
$ bin/highs model.mps

# With options file
$ bin/highs --options_file=my_options.txt model.mps

# Write solution to file
$ bin/highs --solution_file=solution.txt model.mps

# See all command line options
$ bin/highs --help

Example options file (my_options.txt):

solver = pdlp
kkt_tolerance = 1e-4
presolve = on
time_limit = 300

8. C API - Adding Variables and Constraints

// Add a single column (variable)
Highs_addCol(highs, cost, lower, upper, num_new_nz, index, value)

// Add multiple columns
Highs_addCols(highs, num_new_col, costs, lower, upper, num_new_nz,
              starts, index, value)

// Change coefficient in constraint matrix
Highs_changeCoeff(highs, row, col, new_value)

// Change variable bounds
Highs_changeColBounds(highs, col, new_lower, new_upper)

// Change objective coefficient
Highs_changeColCost(highs, col, new_cost)

// Set variable as integer
Highs_changeColIntegrality(highs, col, kHighsVarTypeInteger)

9. Building from Source

# Clone repository
git clone https://github.com/ERGO-Code/HiGHS.git

# Build with CMake
cd HiGHS
cmake -S. -B build
cmake --build build --parallel

# For C# support
cmake -S. -Bbuild -DCSHARP=ON

10. Installation via Package Managers

# Python
$ pip install highspy

# Julia
julia> using Pkg
julia> Pkg.add("HiGHS")

# C# (NuGet)
$ dotnet add package Highs.Native --version 1.12.0

# Linux (install dependencies for HiPO)
$ sudo apt update
$ sudo apt install libopenblas-dev

Reference Files

This skill includes comprehensive documentation in references/:

api.md (15 pages)

Complete API reference for all language interfaces:

• C++ interface: Building from source, class-based API
• Python (highspy): Installation, examples, efficient value extraction
• Julia: JuMP integration, C API wrapper
• C API: Complete function reference for all operations
• C#: NuGet package, build instructions
• Data structures: HighsLp, HighsModel, HighsSparseMatrix, HighsHessian
• Enums: Model status, variable types, solver options

getting_started.md (3 pages)

• Overview of HiGHS capabilities (LP, MILP, QP)
• Installation methods (source, package managers, binaries)
• File format support (MPS, LP, gzip)
• Executable usage and command-line options
• Citation information

guide.md (4 pages)

Advanced features and usage patterns:

• Basic features: Defining models, solving, extracting results
• Further features: Model modification, hot starting, presolve
• GPU acceleration: PDLP solver setup with CUDA
• Feasibility and optimality: Understanding tolerances (absolute vs relative)

options.md (2 pages)

• Complete list of HiGHS options
• How to set options (file, command line, API)
• Important options: presolve, solver, parallel, time_limit, tolerances

solvers.md (1 page)

• Detailed solver descriptions (simplex, IPM, PDLP)
• When to use each solver
• Performance characteristics
• Academic references

terminology.md (1 page)

• Optimization terminology explained
• Bounds, constraints, feasible region
• Sparse matrices
• Primal and dual values
• Basic solutions and sensitivity

Use view to read specific reference files when detailed information is needed.

Key Concepts

Installation

Via Package Managers:

• Python: pip install highspy or conda install highs
• Julia: using Pkg; Pkg.add("HiGHS")
• C#: NuGet package available
• Rust: Available via cargo

From Source:

git clone https://github.com/ERGO-Code/HiGHS.git
cd HiGHS
cmake -S. -B build
cmake --build build --parallel

File Formats Supported

• .mps - MPS format (industry standard)
• .lp - CPLEX LP format
• .gz - Compressed files (gzip)

Data Structures

Key Classes:

• HighsLp - Linear programming model data
• HighsModel - General optimization model (includes QP)
• HighsSparseMatrix - Sparse matrix representation
• HighsHessian - Quadratic objective Hessian
• HighsSolution - Solution data (primal/dual values)
• HighsBasis - Basis status information
• HighsInfo - Solver statistics and convergence info

Important Enums:

• HighsModelStatus - Model status (optimal, infeasible, unbounded, etc.)
• HighsVarType - Variable types (continuous, integer, semi-continuous, etc.)
• ObjSense - Objective sense (minimize, maximize)
• HighsStatus - Return status from API calls

Tolerances and Accuracy

HiGHS uses absolute tolerances by default (default: 1e-7):

• Primal feasibility tolerance - How close constraints must be satisfied
• Dual feasibility tolerance - Dual constraint violations
• Optimality tolerance - KKT condition satisfaction

Important: PDLP uses relative tolerances. For PDLP, increase kkt_tolerance to 1e-4 for faster convergence with acceptable accuracy.

GPU Acceleration

PDLP solver can run on NVIDIA GPUs (Linux/Windows only):

• Requires CUDA Toolkit and matching NVIDIA driver
• Must build HiGHS locally with CMake
• Set solver to "pdlp" and adjust tolerances
• Verify CUDA: nvcc --version

Health Warning: PDLP may not achieve same accuracy as simplex/IPM. Check HighsInfo for actual feasibility values.

Callbacks

HiGHS supports callbacks for:

• Logging custom output
• Implementing custom termination criteria
• Monitoring solver progress
• Extracting intermediate solutions

Working with This Skill

For Beginners

1. Start with references/getting_started.md for installation and basic concepts
2. Review problem formulations in references/terminology.md
3. Try Quick Reference Example 2 (simple Python model)
4. Learn file-based solving with Quick Reference Example 1

For Intermediate Users

• Building models: See examples in references/api.md (Python section)
• Solver options: Read references/options.md for tuning
• Model modification: Check references/guide.md (Further features)
• Hot starting: Use previous solutions to speed up solving

For Advanced Users

• API reference: references/api.md for language-specific details
• Performance tuning: references/solvers.md - choose optimal solver
• GPU acceleration: references/guide.md (GPU section)
• Tolerances: references/guide.md (Feasibility and optimality)

For Code Examples

• Python: references/api.md (Python section) - Complete highspy examples
• Julia: references/api.md (Julia section) - JuMP and C API
• C++: references/api.md (C++ section) - Native library
• C API: references/api.md (C section) - Low-level interface

Performance Benchmarks

HiGHS is competitive with commercial solvers. See:

• Mittelmann LP benchmarks (feasibility and optimality)
• Mittelmann MILP benchmarks

Common Use Cases

1. Supply chain optimization - Minimize costs while meeting constraints
2. Resource allocation - Optimize resource distribution
3. Production planning - Maximize output or minimize waste
4. Portfolio optimization - Balance risk and return (QP)
5. Network flow problems - Transportation, routing
6. Scheduling - Job shop, employee scheduling (MILP)
7. Cutting stock problems - Minimize material waste
8. Energy systems - Power generation and distribution

Resources

references/

Organized documentation extracted from official sources. These files contain:

• Detailed explanations of all features
• Code examples with language annotations
• Links to original documentation
• Table of contents for quick navigation

scripts/

Add helper scripts here for common automation tasks (e.g., batch solving, result analysis).

assets/

Add templates, boilerplate, or example projects here.

Citing HiGHS

If you use HiGHS in an academic context, please cite:

Parallelizing the dual revised simplex method Q. Huangfu and J. A. J. Hall, Mathematical Programming Computation, 10 (1), 119-142, 2018. DOI: 10.1007/s12532-017-0130-5

Troubleshooting

Model Status Not Optimal

• Check HighsInfo for infeasibility/unboundedness indicators
• Review constraint feasibility and objective bounds
• Try different solvers (simplex, IPM, PDLP)
• Verify model data is correct (no NaN, Inf values)

PDLP Reports Optimal But HiGHS Says Not Optimal

• PDLP uses relative tolerances vs absolute
• Increase kkt_tolerance to 1e-4 (recommended)
• Check HighsInfo for actual infeasibility values
• Consider if the solution is "good enough" for your use case

Slow Performance

• Enable presolve: setOptionValue("presolve", "on")
• Try parallel mode for large problems
• Adjust time limits and MIP gap tolerances
• Consider GPU acceleration for very large LPs (PDLP)
• Try different solver: simplex vs IPM vs PDLP

Memory Issues with Large Models

• Use sparse matrix representations
• Pass models via passModel() rather than building incrementally
• Consider model compression techniques
• Use compressed input files (.mps.gz)

Python Performance Issues

• Critical: Convert solution arrays to lists before iteration (see Example 5)
• Use numpy arrays when appropriate
• Avoid repeated API calls in loops

Can't Read LP/MPS File

• Check file format (lpsolve format NOT supported)
• Try compressed version (.gz works, .zip does not)
• Verify file path is correct

Additional Resources

• GitHub: https://github.com/ERGO-Code/HiGHS
• Documentation: https://ergo-code.github.io/HiGHS/dev/
• Contact: highsopt@gmail.com
• Issues: File bugs at GitHub Issues

Notes

• This skill was automatically generated from official HiGHS documentation (v1.10.0+)
• Reference files preserve structure and examples from source docs
• Code examples include language detection for syntax highlighting
• All references link back to original documentation
• Examples tested with HiGHS v1.10.0 and later

Updating

To refresh this skill with updated documentation:

1. Re-run the scraper: python3 cli/doc_scraper.py --config configs/highs.json
2. The skill will be rebuilt with the latest information