Source code for finesse.analysis.actions.optimisation

"""Collection of Actions that deal linear time invariant (LTI) modelling tasks."""

import warnings

from finesse.solutions import BaseSolution
from .base import Action, convert_str_to_parameter
import numpy as np
import logging
from finesse.simulations.sparse.simulation import SparseMatrixSimulation

# from finesse.components.readout import ReadoutDetectorOutput
from finesse.utilities.misc import is_iterable
from finesse.utilities import OrderedSet

LOGGER = logging.getLogger(__name__)


[docs]class OptimizationWarning(RuntimeWarning): pass
[docs]class OptimizeSolution(BaseSolution): """Solution for an optimization action. Attributes ---------- result : scipy.optimize.optimize.OptimizeResult Result from the scipy optimization method that contains the results and some extra data about the process and any errors that might happen. parameters : [Parameter | tuple] Name or names of parameters that were optimized over x : [numeric | ndarray] The final minimized values for the parameters requested. """ def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self.result = None self.x = None self.parameters = None
# IMPORTANT: renaming this class impacts the katscript spec and should be avoided!
[docs]class Optimize(Action): """An action that will optimize the value of `parameter` to either maximize or minimize the output of a `detector` during a simulation. Extra keyword arguments are passed on to the Scipy method: `minimize <https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.minimize.html>`_ This action offers a simplified interface that allows an optimization to be done during a simulation. By default the the Nelder-Mead optimization method is used but can be overriden. The user should read the Scipy documentation to determine which options should be used which are method dependant. Notes ----- Default optimizer used is `nelder-mead`. To set the absolute and relative error targets use (From the scipy documentation: https://docs.scipy.org/doc/scipy/reference/optimize.minimize-neldermead.html) xatol : float, optional Absolute error in xopt between iterations that is acceptable for convergence. Defaults to 1e-4. fatol : float, optional Absolute error in func(xopt) between iterations that is acceptable for convergence. Defaults to 1e-4. These can be set as keyword arguments to the action. Parameters ---------- detector : str The name of the detector output to maximize / minimize. parameter : [:class:`~.Parameter` | str | tuple] The parameter or name of the parameter to optimize, or a tuple of parameters when using multiple targets to optimize over. bounds : list, optional A pair of (lower, upper) bounds on the parameter value. Requires a method that uses bounds. offset : float, optional An offset applied to the detector output when optimizing, defaults to 0. kind : str, optional Either 'max' for maximization or 'min' for minimization, defaults to 'max'. max_iterations : int, optional Maximum number of solver iterations, defaults to 10000. method : str, optional Optimisation method to use, see Scipy documentation for options. name : str, optional The name of this action, defaults to 'maximize'. update_maps : bool, optional If you are changing some parameter or variable that a `Map` depends on then setting this flag to `True` will recompute the `Map` data for each iteration of the optimiser. pre_step : Action, optional Action to run on each step of the optimisation. **kwargs Optional parameters passed to the Scipy optimisation routine as the `options` input. See Scipy method documentation to determine what is available. """ def __init__( self, detector, parameters, bounds=None, offset=0, kind="max", max_iterations=10000, tol=None, verbose=False, method="nelder-mead", opfunc=None, update_maps=False, pre_step=None, name="optimize", **kwargs, ): super().__init__(name) try: self.detector = detector.name except AttributeError: if is_iterable(detector): self.detector = tuple(str(_) for _ in detector) else: self.detector = str(detector) if isinstance(parameters, str): self.parameters = (parameters,) else: try: self.parameters = (*parameters,) except TypeError: self.parameters = (parameters,) self.bounds = bounds self.offset = offset self.kind = kind self.max_iterations = max_iterations self.tol = tol self.kwargs = kwargs self.verbose = verbose self.method = method self.update_maps = update_maps self.opfunc = opfunc self.pre_step = pre_step @property def parameter_names(self): return tuple( param if isinstance(param, str) else param.full_name for param in self.parameters ) def _do(self, state): from scipy.optimize import minimize if state.sim is None: raise Exception("Simulation has not been built") if not isinstance(state.sim, SparseMatrixSimulation): raise NotImplementedError() out_wss = OrderedSet( # workspaces can be in both lists (*state.sim.readout_workspaces, *state.sim.detector_workspaces) ) if not isinstance(self.detector, str): dws = [] for det in self.detector: for ws in out_wss: if ws.oinfo.name == det: dws.append(ws) if len(dws) != len(self.detector): raise RuntimeError( f"Could not find a detector with the name {self.detector}" ) else: dws = None for ws in out_wss: if ws.oinfo.name == self.detector: dws = ws break if dws is None: raise RuntimeError( f"Could not find a detector with the name {self.detector}" ) params = tuple( convert_str_to_parameter(state.model, param) for param in self.parameter_names ) if len(params) == 0: raise RuntimeError( f"Could not find a parameter with the name {self.parameters}" ) sol = OptimizeSolution(self.name) sol.iters = 0 if self.opfunc is None: def func(x, params, state, dws): for a, p in zip(x, params): p.value = a # Determine what the detector workspace needs to calculate # an output. if self.update_maps: state.sim.update_map_data() if self.pre_step: state.apply(self.pre_step) if dws.needs_carrier or dws.needs_signal or dws.needs_noise: state.sim.run_carrier() if dws.needs_signal or dws.needs_noise: state.sim.run_signal(dws.needs_noise) if self.kind == "max": error = -np.abs(np.abs(dws.get_output()) + self.offset) else: error = np.abs(np.abs(dws.get_output()) - self.offset) if self.verbose: print(x, error) return error opfunc = func else: opfunc = self.opfunc res = minimize( opfunc, np.array([_.value for _ in params]), bounds=np.atleast_2d(self.bounds) if self.bounds else None, tol=[self.tol] if self.tol else None, options={"maxiter": self.max_iterations, **self.kwargs}, method=self.method, args=(params, state, dws), ) if res.nit == 1: warnings.warn( "Optimisation finished after 1 iteration, tolerances might be too high", OptimizationWarning, stacklevel=1, ) if self.verbose: if self.offset == 0: print( f"Optimized {self.detector} to {res.fun:.6g} at {self.parameter_names} = {res.x[0]:.6g}" ) else: print( f"Optimized {self.detector} to {self.offset:g}{res.fun:+.6g} at {self.parameter_names} = {res.x[0]:.6g}" ) # in case the values from the last iterations were not the optimal ones # explicitly set the parameter values to the optimal values for opt_value, param in zip(res.x, params): param.value = opt_value sol.result = res sol.x = res.x sol.parameters = self.parameter_names return sol def _requests(self, model, memo, first=True): memo["changing_parameters"].extend(self.parameter_names) # pass along requests needed from action called each step if self.pre_step: self.pre_step._requests(model, memo)
# DDB temp comment out # if self.detector not in model.elements: # raise RuntimeError(f"Could not find a detector called {self.detector}") # det = model.elements[self.detector] # if not (isinstance(det, ReadoutDetectorOutput)) and not ( # np.issubdtype(det.dtype, np.integer) # or np.issubdtype(det.dtype, np.floating) # or np.issubdtype(det.dtype, np.complexfloating) # ): # raise RuntimeError( # f"Detector {self.detector} must output a single integer or floating point value not {det.dtype}" # ) # IMPORTANT: renaming this class impacts the katscript spec and should be avoided!
[docs]class Minimize(Optimize): __doc__ = ( """An action that minimizes some detector output by applying some feedback to multiple targets in a model. """ + Optimize.__doc__ + """ Examples -------- Simple example that minimizes some measured power by feeding back to the laser power:: model = finesse.Model() model.parse(''' l l1 P=1 pd P l1.p1.o ''') sol = model.run("minimize(P, l1.P)") print(sol.result) """ ) def __init__(self, detector, parameter, name="minimize", *args, **kwargs): super().__init__(detector, parameter, *args, name=name, kind="min", **kwargs)
# IMPORTANT: renaming this class impacts the katscript spec and should be avoided!
[docs]class Maximize(Optimize): __doc__ = ( """An action that maximizes some detector output by applying some feedback to multiple targets in a model. """ + Optimize.__doc__ + """ Examples -------- Simple example that maximizes the power in a coupled cavity solution by moving multilpe mirrors:: model = finesse.Model() model.parse(''' l l1 P=1 m m1 R=0.98 T=0.02 phi=10 m m2 R=0.99 T=0.01 m m3 R=1 T=0 phi=-20 link(l1, m1, m2, m3) pd P m3.p1.i ''') sol = model.run("maximize(P, [m1.phi, m3.phi], xatol=1e-7)") print(sol.result) """ ) def __init__(self, detector, parameter, name="maximize", *args, **kwargs): super().__init__(detector, parameter, *args, name=name, kind="max", **kwargs)