[MINIMUM,FVAL,EXITFLAG,OUTPUT] = fminmarkov(FUN,PARS,[OPTIONS],[CONSTRAINTS], ...) Markov Chain Monte Carlo optimizer
This minimization method uses a Markov Chain Monte Carlo, optionally
with constraints on function parameters.
This is an implementation of the Goodman and Weare 2010 Affine
invariant ensemble Markov Chain Monte Carlo (MCMC) sampler. MCMC sampling
enables bayesian inference. The problem with many traditional MCMC samplers
is that they can have slow convergence for badly scaled problems, and that
it is difficult to optimize the random walk for high-dimensional problems.
This is where the GW-algorithm really excels as it is affine invariant. It
can achieve much better convergence on badly scaled problems. It is much
simpler to get to work straight out of the box, and for that reason it
truly deserves to be called the MCMC hammer.
(This code uses a cascaded variant of the Goodman and Weare algorithm).
Calling:
fminmarkov(fun, pars) asks to minimize the 'fun' objective function with starting
parameters 'pars' (vector)
fminmarkov(fun, pars, options) same as above, with customized options (optimset)
fminmarkov(fun, pars, options, fixed)
is used to fix some of the parameters. The 'fixed' vector is then 0 for
free parameters, and 1 otherwise.
fminmarkov(fun, pars, options, lb, ub)
is used to set the minimal and maximal parameter bounds, as vectors.
fminmarkov(fun, pars, options, constraints)
where constraints is a structure (see below).
fminmarkov(problem) where problem is a structure with fields
problem.objective: function to minimize
problem.x0: starting parameter values
problem.options: optimizer options (see below)
problem.constraints: optimization constraints
fminmarkov(..., args, ...)
sends additional arguments to the objective function
criteria = FUN(pars, args, ...)
Example:
banana = @(x)100*(x(2)-x(1)^2)^2+(1-x(1))^2;
[x,fval] = fminmarkov(banana,[-1.2, 1])
Input:
FUN is the function to minimize (handle or string): criteria = FUN(PARS)
It needs to return a single value or vector.
PARS is a vector with initial guess parameters. You must input an
initial guess. PARS can also be given as a single-level structure.
OPTIONS is a structure with settings for the optimizer,
compliant with optimset. Default options may be obtained with
o=fminmarkov('defaults')
options.MinFunEvals sets the minimum number of function evaluations to reach
options.StepSize: unit-less stepsize (default=2.5).
options.ThinChain: Thin all the chains by only storing every N'th step (default=10)
options.BurnIn: fraction of the chain that should be removed. (default=0)
An empty OPTIONS sets the default configuration.
CONSTRAINTS may be specified as a structure
constraints.min= vector of minimal values for parameters
constraints.max= vector of maximal values for parameters
constraints.fixed= vector having 0 where parameters are free, 1 otherwise
constraints.step= vector of maximal parameter changes per iteration
constraints.eval= expression making use of 'p', 'constraints', and 'options'
and returning modified 'p'
or function handle p=@constraints.eval(p)
An empty CONSTRAINTS sets no constraints.
Additional arguments are sent to the objective function.
Output:
MINIMUM is the solution which generated the smallest encountered
value when input into FUN.
FVAL is the value of the FUN function evaluated at MINIMUM.
EXITFLAG return state of the optimizer
OUTPUT additional information returned as a structure.
References:
Goodman & Weare (2010), Ensemble Samplers With Affine Invariance, Comm. App. Math. Comp. Sci., Vol. 5, No. 1, 65-80
Foreman-Mackey, Hogg, Lang, Goodman (2013), emcee: The MCMC Hammer, arXiv:1202.3665
Contrib:
By: Aslak Grinsted 2015 https://github.com/grinsted/gwmcmc
Version: Aug. 22, 2017
See also: fminsearch, optimset
(c) E.Farhi, ILL. License: EUPL.