A simple, (single) header only C++17 library of Nonlinear Numerical Optimizers

The main goal of this library is to be as easy to use as possible, while providing good performance, and enough flexibility to be useful. For this reason, the library has no third party dependencies, requires no complex includes, no additional magic. Just a C++17 compliant compiler and a problem to solve :)

Just copy the header into your project, include and use:

#include "nlsolver.h"

Solvers:

Nelder-Mead
Differential Evolution (see https://ieeexplore.ieee.org/document/5601760 for explanation of different strategies)
- Random Recombination
- Best Recombination
Particle Swarm Optimization
- Vanilla
- Accelerated
Simulated Annealing
- Currently, without option for custom sample generators, only using the Markov Gaussian Kernel
Nelder-Mead PSO hybrid
- Might under-perform other solvers, but should do better than vanilla Nelder-Mead on problems with many local minima where Nelder-Mead can get stuck
- Tends to not be as resource hungry as either of the three previous methods
Gradient Descent
- Several flavours, including Vanilla with fixed steps, optimized using MoreThuente linesearch, and even an implementation of Provably Faster Gradient Descent via Long Steps in case you have a particularly well-behaved smooth convex problem.
BFGS
- fairly standard, but somewhat optimized

Root finding

For one dimensional problems and finding roots of polynomials, the following methods are supported:

Brent's method
Bisection
False Position
ITP
Ridders' method
Tiruneh's method

See roots.cpp or run make roots via the command line for an example across three problems; Generally I would recommend running Brent if you do not really know much about your problem; otherwise ITP can usually be somewhat fast and reliable.

If you have good previous guesses for the root (such as when iteratively optimizing closely related polynomials), Tiruneh's method (see here: https://arxiv.org/abs/1902.09058) is incredibly hard to beat, and very reliable.

Experimental

Levenberg-Marquardt
- semi-experimental; hasn't really been validated beyond a few simple cases

Work in progress

Primarily for hyperparameter optimization,

CMA-ES
Parzen Trees

Roadmap

The goal going forward would be to implement two kinds of constructs for constrained optimization (even though the algorithms implemented here might not be premium solvers for such cases):

Fixed constraints (upper and lower bounds on acceptable parameter values)
Inequality constraints (e.g. $\phi_1 + \phi_2 < 3$) - this will almost surely be implemented as a functor
Even more solvers
Performance benchmarks

Example usage:

Solving the Rosenbrock function

#include "nlsolver.h"

using nlsolver::DESolver;
// random number generators
using nlsolver::rng::xorshift;

class Rosenbrock {
public:
  double operator()(std::vector<double> &x) {
    const double t1 = x[0];
    const double t2 = (x[1] - x[0] * x[0]);
    return t1 * t1 + 100 * t2 * t2;
  }
};

// Different Evolution (DE) solver example
int main() {
  Rosenbrock rosenbrock_prob;
  // DE requires a random number generator - we include some, including a xorshift RNG:
  xorshift<double> gen;
  // initialize solver, supplying objective function and random number generator
  auto de_solver = DESolver<Rosenbrock, xorshift<double>, double> (prob, gen);
  
  // prepare initial values - for DE these work to center the generated agents
  std::vector<double> de_init = {5,7};;
  auto de_res = de_solver.minimize(de_init);
  // print solver status
  de_res.print();
  // best parameters get written back to init
  std::cout << de_init[0] << "," << std::cout << de_init[1] << std::endl;

  return 0;
}

Or run all examples from the command line with make:

make example

And run the examples from the command line

./example

Design notes

There are some design decisions in this library which warrant discussion:

the objective functions to minimize/ maximize are passed as objects, preferring functors, requiring an overloaded public () operator which takes a std::vector, e.g.

struct example {
  double operator()( std::vector<double> &x) {
    return (x[0] + x[1])/(1-x[2]);
  }
};

note that this will also work with lambda functions, and a struct/class is not strictly necessary.[^lambda_note] (See example usage.)

there are no virtual calls in this library - thus incurring no performance penalty
each function exposes both a minimization and a maximization interface, and maximization is always implemented as 'negative minimization', i.e. by multiplying the objective function by -1
all optimizers come with default arguments that try to be sane and user-friendly - but expert users are highly encouraged to supply their own values
currently no multi-threading is supported - this is by design as functors are potentially stateful objective functions and multi-threading would require ensuring no data races happen

Additionally, this library also includes a set of (pseudo)-random and (quasi)-random number generators that also aim to get out of the way as much as possible, all of which are implemented as functors. The library thus assumes that functors are used for random number generation - there is an example on how to use standard library random number generators if one chooses to do so.

Contributing

Feel free to open an issue or create a PR if you feel that an important piece of functionality is missing! Just keep civil, and stay within the spirit of the library (no dependencies, no virtual calls, must support impure objective functions).

Notes

[^lambda_note] This flexibility is included for cases where you want to implicitly bundle mutable data within the struct, and do not want to have to pass the data (e.g. through a pointer) to your objective function. This makes the overall design cleaner - if your objective function needs data, maintains state, or does anything else on evaluation, you can keep the entirety of that within the struct (and even extract it after the solver finishes). If you do not need the functionality, and you simply want to optimize some ad-hoc function, using a lambda is probably much simpler and cleaner.

JSzitas/nlsolver