source: branches/collapse/TEST_OUQ_surrogate_diam_collapse.py @ 855

#!/usr/bin/env python
#
# Author: Lan Huong Nguyen (lanhuong @stanford)
# Copyright (c) 2012-2016 California Institute of Technology.
# License: 3-clause BSD.  The full license text is available at:
#  - http://mmckerns.github.io/project/mystic/browser/mystic/LICENSE

debug = True #False
MINMAX = -1  ## NOTE: sup = maximize = -1; inf = minimize = 1
#######################################################################
# scaling and mpi info; also optimizer configuration parameters
# hard-wired: use DE solver, don't use mpi, F-F' calculation
#######################################################################
npop = 40
maxiter = 6000
maxfun = 1e+6
crossover = 0.9
percent_change = 0.9
convergence_tol = 1e-6
tolW = 0.05; tolP_range_fr = 0.05
ngen = 200; ngcol = 200


#######################################################################
# the model function
#######################################################################
from surrogate import marc_surr as model0

#######################################################################
# the differential evolution optimizer
#######################################################################
def optimize(cost,_bounds,_constraints):
  from mystic.solvers import DifferentialEvolutionSolver2
  from mystic.strategy import Best1Exp
  from mystic.monitors import VerboseMonitor, Monitor
  from mystic.tools import random_seed

  random_seed(122)

  stepmon = VerboseMonitor(10,10)
 #stepmon = Monitor()
  evalmon = Monitor()

  lb,ub,npts = _bounds
  ndim = len(lb)

  solver = DifferentialEvolutionSolver2(ndim,npop)
  solver.SetRandomInitialPoints(min=lb,max=ub)
  solver.SetStrictRanges(min=lb,max=ub)
  solver.SetEvaluationLimits(maxiter,maxfun)
  solver.SetEvaluationMonitor(evalmon)
  solver.SetGenerationMonitor(stepmon)
  solver.SetConstraints(_constraints)
  
  import collapse_code  
  return collapse_code.collapseSolver(solver, cost, npts, lb, ub, convergence_tol, tolW, \
                        tolP_range_fr, ngen, ngcol, Best1Exp, crossover, \
                        percent_change, stepmon, _constraints)
#######################################################################
# maximize the function
#######################################################################
def maximize(params,npts,bounds):

  from mystic.math.measures import split_param
  from mystic.math.discrete import product_measure
  from mystic.math import almostEqual
  from numpy import inf
  atol = 1e-18 # default is 1e-18
  rtol = 1e-7  # default is 1e-7
  target,error = params
  lb,ub = bounds

  # split lower & upper bounds into weight-only & sample-only
  w_lb, x_lb = split_param(lb, npts)
  w_ub, x_ub = split_param(ub, npts)

  # NOTE: rv, lb, ub are of the form:
  #    rv = [wxi]*nx + [xi]*nx + [wyi]*ny + [yi]*ny + [wzi]*nz + [zi]*nz

  # generate primary constraints function
  def constraints(rv, cnstr_x=None):
    c = product_measure()
    c.load(rv, npts)
    # NOTE: bounds wi in [0,1] enforced by filtering
    # impose norm on each discrete measure
    for measure in c:
      if not almostEqual(float(measure.mass), 1.0, tol=atol, rel=rtol):
        measure.normalize()
    # impose expectation on product measure
    ##################### begin function-specific #####################
    E = float(c.expect(model))
    if not (E <= float(target[0] + error[0])) \
    or not (float(target[0] - error[0]) <= E):
      c.set_expect((target[0],error[0]), model, (x_lb,x_ub), cnstr_x)
    ###################### end function-specific ######################
    # extract weights and positions
    return c.flatten()

  # generate maximizing functio
  def cost(rv):
    c = product_measure()
    c.load(rv, npts)
    E = float(c.expect(model))
    if E > (target[0] + error[0]) or E < (target[0] - error[0]):
      if debug: print "skipping expect: %s" % E
      return inf  #XXX: FORCE TO SATISFY E CONSTRAINTS
    return MINMAX * c.pof(model)

  # maximize
  solved, func_max, tot_evaluations, collapse_data = optimize(cost,(lb,ub,npts),constraints)

  if MINMAX == 1:
    print "func_minimum: %s" % func_max  # inf
  else:
    print "func_maximum: %s" % func_max  # sup
  print "tot_evaluations: %s" % tot_evaluations

  return solved, func_max, collapse_data


#######################################################################
# rank, bounds, and restart information 
#######################################################################
if __name__ == '__main__':
  from time import clock
  start = clock()
  
  from mystic.tools import wrap_function
  from mystic.monitors import Null
  fmon = Null()
  model_evaluations, model = wrap_function(model0, [], fmon)  
  function_name = model0.__name__

  H_mean = 6.5    #NOTE: SET THE 'mean' HERE!
  H_range = 1.0   #NOTE: SET THE 'range' HERE!
  nx = 3  #NOTE: SET THE NUMBER OF 'h' POINTS HERE!
  ny = 3  #NOTE: SET THE NUMBER OF 'a' POINTS HERE!
  nz = 3  #NOTE: SET THE NUMBER OF 'v' POINTS HERE!
  target = (H_mean,)
  error = (H_range,)

  w_lower = [0.0]
  w_upper = [1.0]
  h_lower = [60.0];  a_lower = [0.0];  v_lower = [2.1]
  h_upper = [105.0]; a_upper = [30.0]; v_upper = [2.8]

  lower_bounds = (nx * w_lower) + (nx * h_lower) \
               + (ny * w_lower) + (ny * a_lower) \
               + (nz * w_lower) + (nz * v_lower) 
  upper_bounds = (nx * w_upper) + (nx * h_upper) \
               + (ny * w_upper) + (ny * a_upper) \
               + (nz * w_upper) + (nz * v_upper) 

  print "...SETTINGS..."
  print "npop = %s" % npop
  print "ngen = %s" %ngen
  print "ngcol = %s" %ngcol
  print "maxiter = %s" % maxiter
  print "maxfun = %s" % maxfun
  print "convergence_tol = %s" % convergence_tol
  print "crossover = %s" % crossover
  print "percent_change = %s" % percent_change
  print "..............\n"

  print " model: f(x) = %s(x)" % function_name
  print " target: %s" % str(target)
  print " error: %s" % str(error)
  print "..............\n"

  param_string = "["
  for i in range(nx):
    param_string += "'wx%s', " % str(i+1)
  for i in range(nx):
    param_string += "'x%s', " % str(i+1)
  for i in range(ny):
    param_string += "'wy%s', " % str(i+1)
  for i in range(ny):
    param_string += "'y%s', " % str(i+1)
  for i in range(nz):
    param_string += "'wz%s', " % str(i+1)
  for i in range(nz):
    param_string += "'z%s', " % str(i+1)
  param_string = param_string[:-2] + "]"

  print " parameters: %s" % param_string
  print " lower bounds: %s" % lower_bounds
  print " upper bounds: %s" % upper_bounds
# print " ..."
  pars = (target,error)
  npts = (nx,ny,nz)
  bounds = (lower_bounds,upper_bounds)
  solved, diameter, collapse_data = maximize(pars,npts,bounds)

  from numpy import array
  from mystic.math.discrete import product_measure
  c = product_measure()
  c.load(solved,npts)
  print "solved: [wx,x]\n%s" % array(zip(c[0].weights,c[0].positions))
  print "solved: [wy,y]\n%s" % array(zip(c[1].weights,c[1].positions))
  print "solved: [wz,z]\n%s" % array(zip(c[2].weights,c[2].positions))

  print "expect: %s" % str( c.expect(model) )
  
  elapsed = (clock() - start)
  
  print "########################"
  print 'RUNTIME = %s' %elapsed
  print 'COLLAPSE DATA= %s' %collapse_data
  print 'MODEL Evaluations = %s' %model_evaluations
# EOF