Source code for pymanopt.solvers.nelder_mead

import time

import numpy as np

import pymanopt
from pymanopt.solvers.solver import Solver
from pymanopt.solvers.steepest_descent import SteepestDescent


# TODO(nkoep): Check if a suitable autodiff backend is available, and solve the
#              problem using the TR solver if so.
[docs]def compute_centroid(manifold, points): """Compute the centroid of `points` on the `manifold` as Karcher mean.""" num_points = len(points) @pymanopt.function.Callable def objective(y): accumulator = 0 for i in range(num_points): accumulator += manifold.dist(y, points[i]) ** 2 return accumulator / 2 @pymanopt.function.Callable def gradient(y): g = manifold.zerovec(y) for i in range(num_points): g -= manifold.log(y, points[i]) return g # XXX: Manopt runs a few TR iterations here. For us to do this, we either # need to work out the Hessian of the Karcher mean by hand or # implement approximations for the Hessian to use in the TR solver as # Manopt. This is because we cannot implement the Karcher mean with # Theano, say, and compute the Hessian automatically due to dependence # on the manifold-dependent distance function, which is written in # numpy. solver = SteepestDescent(maxiter=15) problem = pymanopt.Problem(manifold, objective, grad=gradient, verbosity=0) return solver.solve(problem)
[docs]class NelderMead(Solver): """ Nelder-Mead minimization alglorithm for derivative-free minimization based on neldermead.m and centroid.m from the manopt MATLAB package. """ def __init__(self, maxcostevals=None, maxiter=None, reflection=1, expansion=2, contraction=0.5, *args, **kwargs): """ Instantiate Nelder-Mead method solver class. Variable attributes (defaults in brackets): - maxcostevals (max(5000, 2 * dim)) Maximum number of allowed cost evaluations - maxiter (max(500, 4 * dim)) Maximum number of allowed iterations - reflection (1) Determines how far to reflect away from the worst vertex; stretched (reflection > 1), compressed (0 < reflection < 1), or exact (reflection = 1) - expansion (2) Factor by which to expand the reflected simplex - contraction (0.5) Factor by which to contract the reflected simplex """ super().__init__(*args, **kwargs) self._maxcostevals = maxcostevals self._maxiter = maxiter self._reflection = reflection self._expansion = expansion self._contraction = contraction
[docs] def solve(self, problem, x=None): """ Perform optimization using a Nelder-Mead minimization algorithm. Arguments: - problem Pymanopt problem setup using the Problem class, this must have a .manifold attribute specifying the manifold to optimize over, as well as a cost and enough information to compute the gradient of that cost. - x=None Optional parameter. Initial population of elements on the manifold. If None then an initial population will be randomly generated Returns: - x Local minimum of obj, or if algorithm terminated before convergence x will be the point at which it terminated """ man = problem.manifold verbosity = problem.verbosity objective = problem.cost # Choose proper default algorithm parameters. We need to know about the # dimension of the manifold to limit the parameter range, so we have to # defer proper initialization until this point. dim = man.dim if self._maxcostevals is None: self._maxcostevals = max(1000, 2 * dim) if self._maxiter is None: self._maxiter = max(2000, 4 * dim) # If no initial simplex x is given by the user, generate one at random. if x is None: x = [man.rand() for i in range(int(dim + 1))] elif not hasattr(x, "__iter__"): raise ValueError("The initial simplex x must be iterable") else: # XXX: Is this necessary? if len(x) != dim + 1: print("The simplex size was adapted to the dimension " "of the manifold") x = x[:dim + 1] # Compute objective-related quantities for x, and setup a function # evaluations counter. costs = np.array([objective(xi) for xi in x]) costevals = dim + 1 # Sort simplex points by cost. order = np.argsort(costs) costs = costs[order] x = [x[i] for i in order] # XXX: Probably inefficient # Iteration counter (at any point, iter is the number of fully executed # iterations so far). iter = 0 time0 = time.time() self._start_optlog() while True: iter += 1 if verbosity >= 2: print("Cost evals: %7d\t" "Best cost: %+.8e" % (costevals, costs[0])) # Sort simplex points by cost. order = np.argsort(costs) costs = costs[order] x = [x[i] for i in order] # XXX: Probably inefficient stop_reason = self._check_stopping_criterion( time0, iter=iter, costevals=costevals) if stop_reason: if verbosity >= 1: print(stop_reason) print('') break # Compute a centroid for the dim best points. xbar = compute_centroid(man, x[:-1]) # Compute the direction for moving along the axis xbar - worst x. vec = man.log(xbar, x[-1]) # Reflection step xr = man.retr(xbar, -self._reflection * vec) costr = objective(xr) costevals += 1 # If the reflected point is honorable, drop the worst point, # replace it by the reflected point and start a new iteration. if costr >= costs[0] and costr < costs[-2]: if verbosity >= 2: print("Reflection") costs[-1] = costr x[-1] = xr continue # If the reflected point is better than the best point, expand. if costr < costs[0]: xe = man.retr(xbar, -self._expansion * vec) coste = objective(xe) costevals += 1 if coste < costr: if verbosity >= 2: print("Expansion") costs[-1] = coste x[-1] = xe continue else: if verbosity >= 2: print("Reflection (failed expansion)") costs[-1] = costr x[-1] = xr continue # If the reflected point is worse than the second to worst point, # contract. if costr >= costs[-2]: if costr < costs[-1]: # do an outside contraction xoc = man.retr(xbar, -self._contraction * vec) costoc = objective(xoc) costevals += 1 if costoc <= costr: if verbosity >= 2: print("Outside contraction") costs[-1] = costoc x[-1] = xoc continue else: # do an inside contraction xic = man.retr(xbar, self._contraction * vec) costic = objective(xic) costevals += 1 if costic <= costs[-1]: if verbosity >= 2: print("Inside contraction") costs[-1] = costic x[-1] = xic continue # If we get here, shrink the simplex around x[0]. if verbosity >= 2: print("Shrinkage") x0 = x[0] for i in np.arange(1, dim + 1): x[i] = man.pairmean(x0, x[i]) costs[i] = objective(x[i]) costevals += dim if self._logverbosity <= 0: return x[0] else: self._stop_optlog(x[0], objective(x[0]), stop_reason, time0, costevals=costevals, iter=iter) return x[0], self._optlog