"""Adaptive Memory Programming for Global Optimization (AMPGO). added to lmfit by Renee Otten (2018) based on the Python implementation of Andrea Gavana (see: http://infinity77.net/global_optimization/) Implementation details can be found in this paper: http://leeds-faculty.colorado.edu/glover/fred%20pubs/416%20-%20AMP%20(TS)%20for%20Constrained%20Global%20Opt%20w%20Lasdon%20et%20al%20.pdf """ import numpy as np from scipy.optimize import minimize SCIPY_LOCAL_SOLVERS = ['Nelder-Mead', 'Powell', 'L-BFGS-B', 'TNC', 'SLSQP'] def ampgo(objfun, x0, args=(), local='L-BFGS-B', local_opts=None, bounds=None, maxfunevals=None, totaliter=20, maxiter=5, glbtol=1e-5, eps1=0.02, eps2=0.1, tabulistsize=5, tabustrategy='farthest', disp=False): """Find the global minimum of a multivariate function using AMPGO. AMPGO stands for "Adaptive Memory Programming for Global Optimization. Parameters ---------- objfun : callable Objective function to be minimized. The function must have the signature:: objfun(params, *args, **kws) x0 : numpy.ndarray Initial guesses for parameter values. args : tuple, optional Additional arguments passed to `objfun`. local : str, optional Name of the local minimization method. Valid options are: - `'L-BFGS-B'` (default) - `'Nelder-Mead'` - `'Powell'` - `'TNC'` - `'SLSQP'` local_opts : dict, optional Options to pass to the local minimizer. bounds : sequence, optional List of tuples specifying the lower and upper bound for each independent variable ``[(xl0, xu0), (xl1, xu1), ...]``. maxfunevals : int, optional Maximum number of function evaluations. If None, the optimization will stop after `totaliter` number of iterations. totaliter : int, optional Maximum number of global iterations. maxiter : int, optional Maximum number of 'Tabu Tunneling' iterations during each global iteration. glbtol : float, optional Tolerance whether or not to accept a solution after a tunneling phase. eps1 : float, optional Constant used to define an aspiration value for the objective function during the Tunneling phase. eps2 : float, optional Perturbation factor used to move away from the latest local minimum at the start of a Tunneling phase. tabulistsize : int, optional Size of the (circular) tabu search list. tabustrategy : {'farthest', 'oldest'}, optional Strategy to use when the size of the tabu list exceeds `tabulistsize`. It can be 'oldest' to drop the oldest point from the tabu list or 'farthest' (default) to drop the element farthest from the last local minimum found. disp : bool, optional Set to True to print convergence messages. Returns ------- tuple A tuple of 5 elements, in the following order: 1. **best_x** (array_like): the estimated position of the global minimum. 2. **best_f** (float): the value of `objfun` at the minimum. 3. **evaluations** (int): the number of function evaluations. 4. **msg** (str): a message describes the cause of the termination. 5. **tunnel_info** (tuple): a tuple containing the total number of Tunneling phases performed and the successful ones. Notes ----- The detailed implementation of AMPGO is described in the paper "Adaptive Memory Programming for Constrained Global Optimization" located here: http://leeds-faculty.colorado.edu/glover/fred%20pubs/416%20-%20AMP%20(TS)%20for%20Constrained%20Global%20Opt%20w%20Lasdon%20et%20al%20.pdf """ if local not in SCIPY_LOCAL_SOLVERS: raise Exception(f'Invalid local solver selected: {local}') x0 = np.atleast_1d(x0) n = len(x0) if bounds is None: bounds = [(None, None)] * n if len(bounds) != n: raise ValueError('length of x0 != length of bounds') bounds = [b if b is not None else (None, None) for b in bounds] _bounds = [(-np.inf if lb is None else lb, np.inf if ub is None else ub) for lb, ub in bounds] low, up = zip(*_bounds) if maxfunevals is None: maxfunevals = np.inf if tabulistsize < 1: raise Exception(f'Invalid tabulistsize specified: {tabulistsize}. ' 'It should be an integer greater than zero.') if tabustrategy not in ['oldest', 'farthest']: raise Exception(f'Invalid tabustrategy specified: {tabustrategy}. ' 'It must be one of "oldest" or "farthest".') tabulist = [] best_f = np.inf best_x = x0 global_iter = 0 all_tunnel = success_tunnel = 0 evaluations = 0 local_tol = min(1e-8, glbtol) while 1: # minimization to find local minimum, either from initial values or # after a successful tunneling loop if disp: print('\n{0}\nStarting MINIMIZATION Phase {1:d}\n{0}' .format('='*72, global_iter+1)) options = {'maxiter': max(1, maxfunevals), 'disp': disp} if local_opts is not None: options.update(local_opts) res = minimize(objfun, x0, args=args, method=local, bounds=bounds, tol=local_tol, options=options) xf, yf, num_fun = res['x'], res['fun'], res['nfev'] if isinstance(yf, np.ndarray): yf = yf[0] maxfunevals -= num_fun evaluations += num_fun if yf < best_f: best_f = yf best_x = xf if disp: print(f'\n\n ==> Reached local minimum: {yf:.5g}\n') if maxfunevals <= 0: if disp: print('='*72) return (best_x, best_f, evaluations, 'Maximum number of function evaluations exceeded', (all_tunnel, success_tunnel)) # if needed, drop a value from the tabu tunneling list and add the # current solution tabulist = drop_tabu_points(xf, tabulist, tabulistsize, tabustrategy) tabulist.append(xf) i = improve = 0 while i < maxiter and improve == 0: if disp: print('{0}\nStarting TUNNELING Phase ({1:d}-{2:d})\n{0}' .format('='*72, global_iter+1, i+1)) all_tunnel += 1 # generate a new starting point away from the current solution r = np.random.uniform(-1.0, 1.0, size=(n, )) beta = eps2*np.linalg.norm(xf) / np.linalg.norm(r) if np.abs(beta) < 1e-8: beta = eps2 x0 = xf + beta*r # make sure that the new starting point is within bounds x0 = np.where(x0 < low, low, x0) x0 = np.where(x0 > up, up, x0) # aspired value of the objective function for the tunneling loop aspiration = best_f - eps1*(1.0 + np.abs(best_f)) tunnel_args = tuple([objfun, aspiration, tabulist] + list(args)) options = {'maxiter': max(1, maxfunevals), 'disp': disp} if local_opts is not None: options.update(local_opts) res = minimize(tunnel, x0, args=tunnel_args, method=local, bounds=bounds, tol=local_tol, options=options) xf, yf, num_fun = res['x'], res['fun'], res['nfev'] if isinstance(yf, np.ndarray): yf = yf[0] maxfunevals -= num_fun evaluations += num_fun yf = inverse_tunnel(xf, yf, aspiration, tabulist) if yf <= best_f + glbtol: oldf = best_f best_f = yf best_x = xf improve = 1 success_tunnel += 1 if disp: print('\n\n ==> Successful tunnelling phase. Reached new ' f'local minimum: {yf:.5g} < {oldf:.5g}\n') i += 1 if maxfunevals <= 0: return (best_x, best_f, evaluations, 'Maximum number of function evaluations exceeded', (all_tunnel, success_tunnel)) tabulist = drop_tabu_points(xf, tabulist, tabulistsize, tabustrategy) tabulist.append(xf) if disp: print('='*72) global_iter += 1 x0 = xf.copy() if global_iter >= totaliter: return (best_x, best_f, evaluations, 'Maximum number of global iterations exceeded', (all_tunnel, success_tunnel)) def drop_tabu_points(xf, tabulist, tabulistsize, tabustrategy): """Drop a point from the tabu search list.""" if len(tabulist) < tabulistsize: return tabulist if tabustrategy == 'oldest': tabulist.pop(0) else: distance = np.sqrt(np.sum((tabulist - xf)**2, axis=1)) index = np.argmax(distance) tabulist.pop(index) return tabulist def tunnel(x0, *args): """Tunneling objective function. This function has a global minimum of zero at any feasible point where ``f(x) = aspiration``, and minimizing this expression tends to move away from all points in `tabulist`. """ objfun, aspiration, tabulist, *fun_args = args numerator = (objfun(x0, *fun_args) - aspiration)**2 denominator = 1.0 for tabu in tabulist: denominator = denominator * np.sqrt(np.sum((x0 - tabu)**2)) ytf = numerator/denominator return ytf def inverse_tunnel(xtf, ytf, aspiration, tabulist): """Calculate the function value after a tunneling phase step.""" denominator = 1.0 for tabu in tabulist: denominator = denominator * np.sqrt(np.sum((xtf - tabu)**2)) numerator = ytf*denominator yf = aspiration + np.sqrt(numerator) return yf