GLP
From OpenOpt
Global Problems (GLP)
-
(global)
- subjected to
-
-
-
-
-
-
Note! Global optimization solvers are slower than NLP/NSP and cannot handle problems with large number of variables (typical sizes are 1..10..100), if you know either you have the only one local=global optimum in your problem, or you search for local one - you'd better use NLP/NSP, they will solve it much faster.
- OpenOpt GLP example
- FuncDesigner GLP examples:
- basic example
- example with guaranteed precision (using interalg)
- (since v. 0.38) another interalg example with discrete variables
GLP solvers connected to OpenOpt:
(most used: interalg, de, pswarm)
| Solver | License | Made by | Are finite box-bounds required | Constraints that can be handled | Info | Parameters |
|---|---|---|---|---|---|---|
| interalg (vectorized) | BSD | Dmitrey | Yes (but they can be very huge) | (since v. 0.36) all | Yields exact optimum subjected to required accuracy fTol: abs(f - f*) < fTol. Can handle categorical variables, general boolean constraints, multiobjective problems. See also interalg benchmark vs Direct, intsolver and commercial BARON. | maxNodes = 15000, maxActiveNodes = 1500, fStart = None, (unestablished) dataHandling = 'auto' | 'raw' | 'sorted' |
| de (vectorized) (works with PyPy since v. >= 0.42 (details) | BSD | Stepan Hlushak, stepanko - at - gmail - dot - com, connected to OO by Dmitrey | Yes | (since v. 0.37) All | Two array differential evolution algorithm (Feoktistov V. Differential Evolution. In Search of Solutions; Springer, 2006, ISBN 0387368965). Code is included into OO and Stepan has subversion commit rights for the one. | baseVectorStrategy = {'random'}/ 'best'; searchDirectionStrategy = {'random'}/'best'; differenceFactorStrategy = {'random'}/'constant'; population = 'default: 10*nVars will be used'; differenceFactor = 0.8; crossoverRate = 0.5; hndvi = 1 (probably hndvi will be changed to more informative name till next OO release), (requires v. >= 0.42) seed = 150880 |
| (requires v. >= 0.42) asa | BSD | Lester Ingber, connected to Python by Robert Jordens, connected to OO by Dmitrey | Yes | lb, ub | Requires pyasa installed (available via easy_install). Algorithm: Adaptive Simulated Annealing. | None yet (could be provided on demand). |
| galileo | GPL | Donald Goodman | Yes | box bounds | GA-based solver. Cannot handle other than box-bound constraints. Code is included into OO. This solver doesn't work with Python 3.x yet. | population = 15; crossoverRate = 1.0; mutationRate = 0.05; useInteger = False (if useInteger = True or 1 then search solution with all integer variables) |
| pswarm (vectorized) | LGPL | A. I. F. Vaz | Seems like no, mb constraints Ax <= b that provide optimization within finite volume are enough | box bounds, linear inequalities | Can handle user-provided x0. Download and install pswarm from the URL mentioned, ensure author-provided RunPSwarm.py works ok, and pswarm_py.so is inside PYTHONPATH. Pay attention: for installation from sources you should use "make py_linear" to enable general linear constraints (Ax <= b). Documentation says pswarm is capable of using parallel calculations (via MPI) but I don't know is it relevant to Python API. The algorithm combines pattern search and particle swarm. Basically, it applies a directional direct search in the poll step (coordinate search in the pure simple bounds case) and particle swarm in the search step. See also: resent paper on PSwarm published at optimization-online.org. | social = 0.5; cognitial = 0.5; fweight = 0.4; iweight = 0.9; size = 42; tol = 1e-5; ddelta = 0.5; idelta = 2.0 |
| stogo | LGPL | Kaj Madsen | Yes | lb, ub | Can use derivatives. Requires nlopt installed. | useRand = True (use GD_STOGO or STOGO_RAND routine, see here for details) |
| isres | LGPL | S. G. Johnson | Yes | All | isres = "Improved Stochastic Ranking Evolution Strategy", by S. G. Johnson. Requires nlopt ver >= 2.2 installed. | population (default 20×(nVars+1)) |
| mlsl | LGPL | S. G. Johnson | Yes | lb, ub | mlsl = "Multi-Level Single-Linkage". This one is for smooth multi-extrema funcs (derivatives are passed to local optimizer). Requires nlopt ver >= 2.2 installed. G_MLSL_LDS is used with LD_TNEWTON_PRECOND_RESTART as local optimizer | population (number of local solvers, default 4) |
Vectorization
(for de and pswarm it works since v. 0.42) Objective function is vectorized for FuncDesigner models and thus the solvers de and pswarm work many times faster (details). As for interalg, it is also vectorized (in a somewhat different way although) but first of all it's speed depends on quality of interval analysis for objective and constraints.
See also:
- StochasticProgramming - commercial addon for FuncDesigner, free for small-scaled research / educational purposes
- Why SciPy anneal is not included
- MOP (multi-objective problems)
- NLP (nonlinear problems)
- NSP (nonsmooth problems)
- Global Optimization portal by Arnold Neumaier
