[3570545] | 1 | #__docformat__ = "restructuredtext en" |
---|
| 2 | # ******NOTICE*************** |
---|
| 3 | # from optimize.py module by Travis E. Oliphant |
---|
| 4 | # |
---|
| 5 | # You may copy and use this module as you see fit with no |
---|
| 6 | # guarantee implied provided you keep this notice in all copies. |
---|
| 7 | # *****END NOTICE************ |
---|
| 8 | # |
---|
| 9 | # Modified by Paul Kienzle to support bounded minimization |
---|
| 10 | """ |
---|
| 11 | Downhill simplex optimizer. |
---|
| 12 | """ |
---|
| 13 | |
---|
| 14 | __all__ = ['simplex'] |
---|
| 15 | |
---|
| 16 | __docformat__ = "restructuredtext en" |
---|
| 17 | |
---|
| 18 | import numpy |
---|
| 19 | __version__="0.7" |
---|
| 20 | |
---|
| 21 | def wrap_function(function, bounds): |
---|
| 22 | ncalls = [0] |
---|
| 23 | if bounds is not None: |
---|
| 24 | lo, hi = [numpy.asarray(v) for v in bounds] |
---|
| 25 | def function_wrapper(x): |
---|
| 26 | ncalls[0] += 1 |
---|
| 27 | if numpy.any((x<lo)|(x>hi)): |
---|
| 28 | return numpy.inf |
---|
| 29 | else: |
---|
| 30 | return function(x) |
---|
| 31 | else: |
---|
| 32 | def function_wrapper(x): |
---|
| 33 | ncalls[0] += 1 |
---|
| 34 | return function(x) |
---|
| 35 | return ncalls, function_wrapper |
---|
| 36 | |
---|
| 37 | class Result: |
---|
| 38 | """ |
---|
| 39 | Results from the fit. |
---|
| 40 | |
---|
| 41 | x : ndarray |
---|
| 42 | Best parameter set |
---|
| 43 | fx : float |
---|
| 44 | Best value |
---|
| 45 | iters : int |
---|
| 46 | Number of iterations |
---|
| 47 | calls : int |
---|
| 48 | Number of function calls |
---|
| 49 | success : boolean |
---|
| 50 | True if the fit completed successful, false if terminated early |
---|
| 51 | because of too many iterations. |
---|
| 52 | """ |
---|
| 53 | def __init__(self, x, fx, iters, calls, status): |
---|
| 54 | self.x,self.fx,self.iters,self.calls=x,fx,iters,calls |
---|
| 55 | self.status = status |
---|
| 56 | def __str__(self): |
---|
| 57 | return "Minimum %g at %s after %d calls"%(self.fx,self.x,self.calls) |
---|
| 58 | |
---|
| 59 | def dont_abort(): return False |
---|
| 60 | |
---|
| 61 | def simplex(f, x0=None, bounds=None, radius=0.05, |
---|
| 62 | xtol=1e-4, ftol=1e-4, maxiter=None, |
---|
| 63 | update_handler=None, abort_test=dont_abort): |
---|
| 64 | """ |
---|
| 65 | Minimize a function using Nelder-Mead downhill simplex algorithm. |
---|
| 66 | |
---|
| 67 | This optimizer is also known as Amoeba (from Numerical Recipes) and |
---|
| 68 | the Nealder-Mead simplex algorithm. This is not the simplex algorithm |
---|
| 69 | for solving constrained linear systems. |
---|
| 70 | |
---|
| 71 | Downhill simplex is a robust derivative free algorithm for finding |
---|
| 72 | minima. It proceeds by choosing a set of points (the simplex) forming |
---|
| 73 | an n-dimensional triangle, and transforming that triangle so that the |
---|
| 74 | worst vertex is improved, either by stretching, shrinking or reflecting |
---|
| 75 | it about the center of the triangle. This algorithm is not known for |
---|
| 76 | its speed, but for its simplicity and robustness, and is a good algorithm |
---|
| 77 | to start your problem with. |
---|
| 78 | |
---|
| 79 | *Parameters*: |
---|
| 80 | |
---|
| 81 | f : callable f(x,*args) |
---|
| 82 | The objective function to be minimized. |
---|
| 83 | x0 : ndarray |
---|
| 84 | Initial guess. |
---|
| 85 | bounds : (ndarray,ndarray) or None |
---|
| 86 | Bounds on the parameter values for the function. |
---|
| 87 | radius: float |
---|
| 88 | Size of the initial simplex. For bounded parameters (those |
---|
| 89 | which have finite lower and upper bounds), radius is clipped |
---|
| 90 | to a value in (0,0.5] representing the portion of the |
---|
| 91 | range to use as the size of the initial simplex. |
---|
| 92 | |
---|
| 93 | *Returns*: Result (`park.simplex.Result`) |
---|
| 94 | |
---|
| 95 | x : ndarray |
---|
| 96 | Parameter that minimizes function. |
---|
| 97 | fx : float |
---|
| 98 | Value of function at minimum: ``fopt = func(xopt)``. |
---|
| 99 | iters : int |
---|
| 100 | Number of iterations performed. |
---|
| 101 | calls : int |
---|
| 102 | Number of function calls made. |
---|
| 103 | success : boolean |
---|
| 104 | True if fit completed successfully. |
---|
| 105 | |
---|
| 106 | *Other Parameters*: |
---|
| 107 | |
---|
| 108 | xtol : float |
---|
| 109 | Relative error in xopt acceptable for convergence. |
---|
| 110 | ftol : number |
---|
| 111 | Relative error in func(xopt) acceptable for convergence. |
---|
| 112 | maxiter : int=200*N |
---|
| 113 | Maximum number of iterations to perform. Defaults |
---|
| 114 | update_handler : callable |
---|
| 115 | Called after each iteration, as callback(xk,fxk), where xk |
---|
| 116 | is the current parameter vector and fxk is the function value. |
---|
| 117 | Returns True if the fit should continue. |
---|
| 118 | |
---|
| 119 | |
---|
| 120 | *Notes* |
---|
| 121 | |
---|
| 122 | Uses a Nelder-Mead simplex algorithm to find the minimum of |
---|
| 123 | function of one or more variables. |
---|
| 124 | |
---|
| 125 | """ |
---|
| 126 | fcalls, func = wrap_function(f, bounds) |
---|
| 127 | x0 = numpy.asfarray(x0).flatten() |
---|
| 128 | #print "x0",x0 |
---|
| 129 | N = len(x0) |
---|
| 130 | rank = len(x0.shape) |
---|
| 131 | if not -1 < rank < 2: |
---|
| 132 | raise ValueError, "Initial guess must be a scalar or rank-1 sequence." |
---|
| 133 | |
---|
| 134 | if maxiter is None: |
---|
| 135 | maxiter = N * 200 |
---|
| 136 | |
---|
| 137 | rho = 1; chi = 2; psi = 0.5; sigma = 0.5; |
---|
| 138 | |
---|
| 139 | if rank == 0: |
---|
| 140 | sim = numpy.zeros((N+1,), dtype=x0.dtype) |
---|
| 141 | else: |
---|
| 142 | sim = numpy.zeros((N+1,N), dtype=x0.dtype) |
---|
| 143 | fsim = numpy.zeros((N+1,), float) |
---|
| 144 | sim[0] = x0 |
---|
| 145 | fsim[0] = func(x0) |
---|
| 146 | |
---|
| 147 | # Metropolitan simplex: simplex has vertices at x0 and at |
---|
| 148 | # x0 + j*radius for each unit vector j. Radius is a percentage |
---|
| 149 | # change from the initial value, or just the radius if the initial |
---|
| 150 | # value is 0. For bounded problems, the radius is a percentage of |
---|
| 151 | # the bounded range in dimension j. |
---|
| 152 | val = x0*(1+radius) |
---|
| 153 | val[val == 0] = radius |
---|
| 154 | if bounds is not None: |
---|
| 155 | radius = numpy.clip(radius,0,0.5) |
---|
| 156 | lo,hi = [numpy.asarray(v) for v in bounds] |
---|
| 157 | |
---|
| 158 | # Keep the initial simplex inside the bounds |
---|
[fc049b7] | 159 | x0[x0<lo] = lo[x0<lo] |
---|
| 160 | x0[x0>hi] = hi[x0>hi] |
---|
[3570545] | 161 | bounded = ~numpy.isinf(lo) & ~numpy.isinf(hi) |
---|
| 162 | val[bounded] = x0[bounded] + (hi[bounded]-lo[bounded])*radius |
---|
[fc049b7] | 163 | val[val<lo] = lo[val<lo] |
---|
| 164 | val[val>hi] = hi[val>hi] |
---|
[3570545] | 165 | |
---|
| 166 | # If the initial point was at or beyond an upper bound, then bounds |
---|
| 167 | # projection will put x0 and x0+j*radius at the same point. We |
---|
| 168 | # need to detect these collisions and reverse the radius step |
---|
| 169 | # direction when such collisions occur. The only time the collision |
---|
| 170 | # can occur at the lower bound is when upper and lower bound are |
---|
| 171 | # identical. In that case, we are already done. |
---|
| 172 | collision = val==x0 |
---|
| 173 | if numpy.any(collision): |
---|
| 174 | reverse = x0*(1-radius) |
---|
| 175 | reverse[reverse==0] = -radius |
---|
| 176 | reverse[bounded] = x0[bounded] - (hi[bounded]-lo[bounded])*radius |
---|
| 177 | val[collision] = reverse[collision] |
---|
| 178 | |
---|
| 179 | # Make tolerance relative for bounded parameters |
---|
| 180 | tol = numpy.ones(x0.shape)*xtol |
---|
| 181 | tol[bounded] = (hi[bounded]-lo[bounded])*xtol |
---|
| 182 | xtol = tol |
---|
| 183 | |
---|
| 184 | # Compute values at the simplex vertices |
---|
| 185 | for k in range(0,N): |
---|
| 186 | y = x0+0 |
---|
| 187 | y[k] = val[k] |
---|
| 188 | sim[k+1] = y |
---|
| 189 | fsim[k+1] = func(y) |
---|
| 190 | |
---|
| 191 | #print sim |
---|
| 192 | ind = numpy.argsort(fsim) |
---|
| 193 | fsim = numpy.take(fsim,ind,0) |
---|
| 194 | # sort so sim[0,:] has the lowest function value |
---|
| 195 | sim = numpy.take(sim,ind,0) |
---|
| 196 | #print sim |
---|
| 197 | |
---|
| 198 | iterations = 1 |
---|
| 199 | while iterations < maxiter: |
---|
| 200 | if numpy.all(abs(sim[1:]-sim[0]) <= xtol) \ |
---|
| 201 | and max(abs(fsim[0]-fsim[1:])) <= ftol: |
---|
| 202 | #print abs(sim[1:]-sim[0]) |
---|
| 203 | break |
---|
| 204 | |
---|
| 205 | xbar = numpy.sum(sim[:-1],0) / N |
---|
| 206 | xr = (1+rho)*xbar - rho*sim[-1] |
---|
| 207 | #print "xbar" ,xbar,rho,sim[-1],N |
---|
| 208 | #break |
---|
| 209 | fxr = func(xr) |
---|
| 210 | doshrink = 0 |
---|
| 211 | |
---|
| 212 | if fxr < fsim[0]: |
---|
| 213 | xe = (1+rho*chi)*xbar - rho*chi*sim[-1] |
---|
| 214 | fxe = func(xe) |
---|
| 215 | |
---|
| 216 | if fxe < fxr: |
---|
| 217 | sim[-1] = xe |
---|
| 218 | fsim[-1] = fxe |
---|
| 219 | else: |
---|
| 220 | sim[-1] = xr |
---|
| 221 | fsim[-1] = fxr |
---|
| 222 | else: # fsim[0] <= fxr |
---|
| 223 | if fxr < fsim[-2]: |
---|
| 224 | sim[-1] = xr |
---|
| 225 | fsim[-1] = fxr |
---|
| 226 | else: # fxr >= fsim[-2] |
---|
| 227 | # Perform contraction |
---|
| 228 | if fxr < fsim[-1]: |
---|
| 229 | xc = (1+psi*rho)*xbar - psi*rho*sim[-1] |
---|
| 230 | fxc = func(xc) |
---|
| 231 | |
---|
| 232 | if fxc <= fxr: |
---|
| 233 | sim[-1] = xc |
---|
| 234 | fsim[-1] = fxc |
---|
| 235 | else: |
---|
| 236 | doshrink=1 |
---|
| 237 | else: |
---|
| 238 | # Perform an inside contraction |
---|
| 239 | xcc = (1-psi)*xbar + psi*sim[-1] |
---|
| 240 | fxcc = func(xcc) |
---|
| 241 | |
---|
| 242 | if fxcc < fsim[-1]: |
---|
| 243 | sim[-1] = xcc |
---|
| 244 | fsim[-1] = fxcc |
---|
| 245 | else: |
---|
| 246 | doshrink = 1 |
---|
| 247 | |
---|
| 248 | if doshrink: |
---|
| 249 | for j in xrange(1,N+1): |
---|
| 250 | sim[j] = sim[0] + sigma*(sim[j] - sim[0]) |
---|
| 251 | fsim[j] = func(sim[j]) |
---|
| 252 | |
---|
| 253 | ind = numpy.argsort(fsim) |
---|
| 254 | sim = numpy.take(sim,ind,0) |
---|
| 255 | fsim = numpy.take(fsim,ind,0) |
---|
| 256 | if update_handler is not None: |
---|
| 257 | update_handler(sim[0],fsim[0]) |
---|
| 258 | iterations += 1 |
---|
| 259 | if abort_test(): break |
---|
| 260 | |
---|
| 261 | status = 0 if iterations < maxiter else 1 |
---|
| 262 | res = Result(sim[0],fsim[0],iterations,fcalls[0], status) |
---|
| 263 | return res |
---|
| 264 | |
---|
| 265 | def main(): |
---|
| 266 | import time |
---|
| 267 | def rosen(x): # The Rosenbrock function |
---|
| 268 | return numpy.sum(100.0*(x[1:]-x[:-1]**2.0)**2.0 + (1-x[:-1])**2.0,axis=0) |
---|
| 269 | |
---|
| 270 | |
---|
| 271 | x0 = [0.8,1.2,0.7] |
---|
| 272 | print "Nelder-Mead Simplex" |
---|
| 273 | print "===================" |
---|
| 274 | start = time.time() |
---|
| 275 | x = simplex(rosen,x0) |
---|
| 276 | print x |
---|
| 277 | print "Time:",time.time() - start |
---|
| 278 | |
---|
| 279 | x0 = [0]*3 |
---|
| 280 | print "Nelder-Mead Simplex" |
---|
| 281 | print "===================" |
---|
| 282 | print "starting at zero" |
---|
| 283 | start = time.time() |
---|
| 284 | x = simplex(rosen,x0) |
---|
| 285 | print x |
---|
| 286 | print "Time:",time.time() - start |
---|
| 287 | |
---|
| 288 | x0 = [0.8,1.2,0.7] |
---|
| 289 | lo,hi = [0]*3, [1]*3 |
---|
| 290 | print "Bounded Nelder-Mead Simplex" |
---|
| 291 | print "===========================" |
---|
| 292 | start = time.time() |
---|
| 293 | x = simplex(rosen,x0,bounds=(lo,hi)) |
---|
| 294 | print x |
---|
| 295 | print "Time:",time.time() - start |
---|
| 296 | |
---|
| 297 | |
---|
| 298 | x0 = [0.8,1.2,0.7] |
---|
| 299 | lo,hi = [0.999]*3, [1.001]*3 |
---|
| 300 | print "Bounded Nelder-Mead Simplex" |
---|
| 301 | print "===========================" |
---|
| 302 | print "tight bounds" |
---|
| 303 | print "simplex is smaller than 1e-7 in every dimension, but you can't" |
---|
| 304 | print "see this without uncommenting the print statement simplex function" |
---|
| 305 | start = time.time() |
---|
| 306 | x = simplex(rosen,x0,bounds=(lo,hi),xtol=1e-4) |
---|
| 307 | print x |
---|
| 308 | print "Time:",time.time() - start |
---|
| 309 | |
---|
| 310 | |
---|
| 311 | x0 = [0]*3 |
---|
| 312 | hi,lo = [-0.999]*3, [-1.001]*3 |
---|
| 313 | print "Bounded Nelder-Mead Simplex" |
---|
| 314 | print "===========================" |
---|
| 315 | print "tight bounds, x0=0 outside bounds from above" |
---|
| 316 | start = time.time() |
---|
| 317 | x = simplex(lambda x:rosen(-x),x0,bounds=(lo,hi),xtol=1e-4) |
---|
| 318 | print x |
---|
| 319 | print "Time:",time.time() - start |
---|
| 320 | |
---|
| 321 | |
---|
| 322 | x0 = [0.8,1.2,0.7] |
---|
| 323 | lo,hi = [-numpy.inf]*3, [numpy.inf]*3 |
---|
| 324 | print "Bounded Nelder-Mead Simplex" |
---|
| 325 | print "===========================" |
---|
| 326 | print "infinite bounds" |
---|
| 327 | start = time.time() |
---|
| 328 | x = simplex(rosen,x0,bounds=(lo,hi),xtol=1e-4) |
---|
| 329 | print x |
---|
| 330 | print "Time:",time.time() - start |
---|
| 331 | |
---|
| 332 | if __name__ == "__main__": |
---|
| 333 | main() |
---|