13.15. Calculation algorithm “SimulatedAnnealing”¶

13.15.1. Description¶

This algorithm estimates the state of a system by gradient-free minimization of a cost function $J$ , using the simulated annealing search meta-heuristic. It is based on reducing the $J$ error as far as possible, while allowing a temporary rise in this error to avoid getting stuck in a local minimum. The rise in error is driven by a statistical law of temperature, hence the analogy with metal annealing, which gives the method its name. The method does not require any particular information on the functional, nor does it need derivatives (except in its hybrid “DualAnnealing” version).

It falls in the same category than the Calculation algorithm “DerivativeFreeOptimization”, Calculation algorithm “DifferentialEvolution”, Calculation algorithm “ParticleSwarmOptimization”, Calculation algorithm “TabuSearch”.

It is a single-objective optimization method, allowing the search for the global minimum of any cost function $J$ of type $L^1$ , $L^2$ or $L^{\infty}$ , with or without weights, as described in the section for Going further in the state estimation by optimization methods. As it is a meta-heuristic, except in special cases, reaching a global or local optimal result is not guaranteed (except in its hybrid “DualAnnealing” version). The default error functional is that of weighted augmented least squares, classically used in data assimilation.

There exists various variants of this algorithm. The following stable and robust formulations are proposed here:

“GeneralizedSimulatedAnnealing” (Generalized Simulated Annealing or GSA, see [Tsallis96]), a classical algorithm combining classical and fast simulated annealing approaches. It is powerful, robust and represents a reference for simulated annealing methods.
“DualAnnealing” (see [Xiang97]), an algorithm combining the previous GSA with a local search strategy, applied to states acceptable from the point of view of simulated annealing, which improves the speed and accuracy of the GSA. This improvement requires the tangent and adjoint operators.

The following are a few practical suggestions for the effective use of these algorithms:

The recommended variant of this algorithm is “DualAnnealing”, as it is both robust and converges very well, especially in high dimensions for such an algorithm. However, as it requires the tangent and adjoint operators, it is not always wise to use this acceleration feature.
In cases where the error function or the observation operator are not derivable, the “GeneralizedSimulatedAnnealing” algorithm is suitable and achieves the same simulated annealing optimization as the accelerated variant.
The easiest way to check convergence is to leave the parameters at their default values and let the simulated annealing stabilize. However, since stochastic convergence can take a long time, it is also possible to restrict calculations by limiting the number of simulation function evaluations. This does not limit theoretical convergence, but it does significantly reduce the number of calculations.

These suggestions are to be used as experimental indications, not as requirements, because they are to be appreciated or adapted according to the physics of each problem that is treated.

13.15.2. Some noteworthy properties of the implemented methods¶

To complete the description, we summarize here a few notable properties of the algorithm methods or of their implementations. These properties may have an influence on how it is used or on its computational performance. For further information, please refer to the more comprehensive references given at the end of this algorithm description.

The optimization methods proposed by this algorithm perform a global search for the minimum, theoretically achieving a globally optimal state over the search domain. However, this global optimality is achieved “at convergence”, which means in long or infinite time during an iterative optimization “with real values” (as opposed to “with integer values”).

The methods proposed by this algorithm do not require derivation of the objective function or of one of the operators, thus avoiding this additional calculation time when derivatives are calculated numerically by multiple evaluations.

The methods proposed by this algorithm have no internal parallelism or numerical derivation of operator(s), and therefore cannot take advantage of computer resources for distributing calculations. The methods are sequential, and any use of parallelism resources is therefore reserved for observation or evolution operators, i.e. user codes.

The methods proposed by this algorithm achieve their convergence on one or more residue or number criteria. In practice, there may be several convergence criteria active simultaneously.

The residue can be a conventional measure based on a gap (e.g. “calculation-measurement gap”), or be a significant value for the algorithm (e.g. “nullity of gradient”).

The number is frequently a significant value for the algorithm, such as a number of iterations or a number of evaluations, but it can also be, for example, a number of generations for an evolutionary algorithm.

Convergence thresholds need to be carefully adjusted, to reduce the gobal calculation cost, or to ensure that convergence is adapted to the physical case encountered.

13.15.3. Optional and required commands¶

The general required commands, available in the editing user graphical or textual interface, are the following:

Background: Vector. The variable indicates the background or initial vector used, previously noted as $\mathbf{x}^b$ . Its value is defined as a “Vector” or “VectorSerie” type object. Its availability in output is conditioned by the boolean “Stored” associated with input.

BackgroundError: Matrix. This indicates the background error covariance matrix, previously noted as $\mathbf{B}$ . Its value is defined as a “Matrix” type object, a “ScalarSparseMatrix” type object, or a “DiagonalSparseMatrix” type object, as described in detail in the section Requirements to describe covariance matrices. Its availability in output is conditioned by the boolean “Stored” associated with input.

EvolutionError: Matrix. The variable indicates the evolution error covariance matrix, usually noted as $\mathbf{Q}$ . It is defined as a “Matrix” type object, a “ScalarSparseMatrix” type object, or a “DiagonalSparseMatrix” type object, as described in detail in the section Requirements to describe covariance matrices. Its availability in output is conditioned by the boolean “Stored” associated with input.

EvolutionModel: Operator. The variable indicates the evolution model operator, usually noted $M$ , which describes an elementary step of evolution. Its value is defined as a “Function” type object or a “Matrix” type one. In the case of “Function” type, different functional forms can be used, as described in the section Requirements for functions describing an operator. If there is some control $U$ included in the evolution model, the operator has to be applied to a pair $(X,U)$ .

Observation: List of vectors. The variable indicates the observation vector used for data assimilation or optimization, and usually noted $\mathbf{y}^o$ . Its value is defined as an object of type “Vector” if it is a single observation (temporal or not) or “VectorSeries” if it is a succession of observations. Its availability in output is conditioned by the boolean “Stored” associated in input.

ObservationError: Matrix. The variable indicates the observation error covariance matrix, usually noted as $\mathbf{R}$ . It is defined as a “Matrix” type object, a “ScalarSparseMatrix” type object, or a “DiagonalSparseMatrix” type object, as described in detail in the section Requirements to describe covariance matrices. Its availability in output is conditioned by the boolean “Stored” associated with input.

ObservationOperator: Operator. The variable indicates the observation operator, usually noted as $H$ , which transforms the input parameters $\mathbf{x}$ to results $\mathbf{y}$ to be compared to observations $\mathbf{y}^o$ . Its value is defined as a “Function” type object or a “Matrix” type one. In the case of “Function” type, different functional forms can be used, as described in the section Requirements for functions describing an operator. If there is some control $U$ included in the observation, the operator has to be applied to a pair $(X,U)$ .

The general optional commands, available in the editing user graphical or textual interface, are indicated in List of commands and keywords for data assimilation or optimization case. Moreover, the parameters of the command “AlgorithmParameters” allows to choose the specific options, described hereafter, of the algorithm. See Description of options of an algorithm by “AlgorithmParameters” for the good use of this command.

The options are the following:

Bounds

List of pairs of real values. This key allows to define pairs of upper and lower bounds for every state variable being optimized. Bounds have to be given by a list of list of pairs of lower/upper bounds for each variable, with a value of None each time there is no bound. The bounds can always be specified, but they are taken into account only by the constrained optimizers. If the list is empty, there are no bounds.

Example: {"Bounds":[[2.,5.],[1.e-2,10.],[-30.,None],[None,None]]}

EstimationOf

Predefined name. This key allows to choose the type of estimation to be performed. It can be either state-estimation, with a value of “State”, or parameter-estimation, with a value of “Parameters”. The default choice is “Parameters”.

Example: {"EstimationOf":"Parameters"}

MaximumNumberOfIterations

Integer value. This key indicates the maximum number of internal iterations allowed for iterative optimization. The default is 15000, which is very similar to no limit on iterations. It is then recommended to adapt this parameter to the needs on real problems. For some optimizers, the effective stopping step can be slightly different of the limit due to algorithm internal control requirements. One can refer to the section describing ways for Convergence control for calculation cases and iterative algorithms for more detailed recommendations.

Example: {"MaximumNumberOfIterations":100}

MaximumNumberOfFunctionEvaluations

Integer value. This key indicates the maximum number of evaluation of the cost function to be optimized. The default is 15000, which is an arbitrary limit. It is then recommended to adapt this parameter to the needs on real problems. For some optimizers, the effective number of function evaluations can be slightly different of the limit due to algorithm internal control requirements.

Example: {"MaximumNumberOfFunctionEvaluations":50}

Minimizer: Predefined name. This key allows to choose the optimization minimizer. The default choice is “LBFGSB”, and the possible ones are “LBFGSB” (nonlinear constrained minimizer, see [Byrd95], [Morales11], [Zhu97]), “TNC” (nonlinear constrained minimizer), “CG” (nonlinear unconstrained minimizer), “BFGS” (nonlinear unconstrained minimizer), It is highly recommended to keep the default value.

QualityCriterion

Predefined name. This key indicates the quality criterion, minimized to find the optimal state estimate. The default is the usual data assimilation criterion named “DA”, the augmented weighted least squares. The possible criterion has to be in the following list, where the equivalent names are indicated by the sign “<=>”: [“AugmentedWeightedLeastSquares” <=> “AWLS” <=> “DA”, “WeightedLeastSquares” <=> “WLS”, “LeastSquares” <=> “LS” <=> “L2”, “AbsoluteValue” <=> “L1”, “MaximumError” <=> “ME” <=> “Linf”]. See the section for Going further in the state estimation by optimization methods to have a detailed definition of these quality criteria.

Example: {"QualityCriterion":"DA"}

SetSeed

Integer value. This key allow to give an integer in order to fix the seed of the random generator used in the algorithm. By default, the seed is left uninitialized, and so use the default initialization from the computer, which then change at each study. To ensure the reproducibility of results involving random samples, it is strongly advised to initialize the seed. A simple convenient value is for example 123456789. It is recommended to put an integer with more than 6 or 7 digits to properly initialize the random generator.

Example: {"SetSeed":123456789}

StoreSupplementaryCalculations

List of names. This list indicates the names of the supplementary variables, that can be available during or at the end of the algorithm, if they are initially required by the user. Their availability involves, potentially, costly calculations or memory consumptions. The default is then a void list, none of these variables being calculated and stored by default (excepted the unconditional variables). The possible names are in the following list (the detailed description of each named variable is given in the following part of this specific algorithmic documentation, in the sub-section “Information and variables available at the end of the algorithm”): [ “Analysis”, “BMA”, “CostFunctionJ”, “CostFunctionJb”, “CostFunctionJo”, “CostFunctionJAtCurrentOptimum”, “CostFunctionJbAtCurrentOptimum”, “CostFunctionJoAtCurrentOptimum”, “CurrentIterationNumber”, “CurrentOptimum”, “CurrentState”, “EnsembleOfSimulations”, “EnsembleOfStates”, “IndexOfOptimum”, “Innovation”, “InnovationAtCurrentState”, “OMA”, “OMB”, “SimulatedObservationAtBackground”, “SimulatedObservationAtCurrentOptimum”, “SimulatedObservationAtCurrentState”, “SimulatedObservationAtOptimum”, ].

Example : {"StoreSupplementaryCalculations":["CurrentState", "Residu"]}

Variant

Predefined name. This key allows to choose one of the possible variants for the main algorithm. The default variant is the original “DualAnnealing”, and the possible choices are “GeneralizedSimulatedAnnealing” (Generalized Simulated Annealing ou GSA), “DualAnnealing” (Dual Annealing).

It is recommended to try the “DualAnnealing” variant with a very small number of iterations and a limited number of simulation function evaluations.

Example : {"Variant":"DualAnnealing"}

13.15.4. Information and variables available at the end of the algorithm¶

At the output, after executing the algorithm, there are information and variables originating from the calculation. The description of Variables and information available at the output show the way to obtain them by the method named get, of the variable “ADD” of the post-processing in graphical interface, or of the case in textual interface. The input variables, available to the user at the output in order to facilitate the writing of post-processing procedures, are described in an Inventory of potentially available information at the output.

Permanent outputs (non conditional)

The unconditional outputs of the algorithm are the following:

Analysis

List of vectors. Each element of this variable is an optimal state $\mathbf{x}^*$ in optimization, an interpolate or an analysis $\mathbf{x}^a$ in data assimilation.

Example: xa = ADD.get("Analysis")[-1]

CostFunctionJ

List of values. Each element is a value of the chosen error function $J$ .

Example: J = ADD.get("CostFunctionJ")[:]

CostFunctionJb

List of values. Each element is a value of the error function $J^b$ , that is of the background difference part. If this part does not exist in the error function, its value is zero.

Example: Jb = ADD.get("CostFunctionJb")[:]

CostFunctionJo

List of values. Each element is a value of the error function $J^o$ , that is of the observation difference part.

Example: Jo = ADD.get("CostFunctionJo")[:]

Set of on-demand outputs (conditional or not)

The whole set of algorithm outputs (conditional or not), sorted by alphabetical order, is the following:

Analysis

List of vectors. Each element of this variable is an optimal state $\mathbf{x}^*$ in optimization, an interpolate or an analysis $\mathbf{x}^a$ in data assimilation.

Example: xa = ADD.get("Analysis")[-1]

BMA

List of vectors. Each element is a vector of difference between the background and the optimal state.

Example: bma = ADD.get("BMA")[-1]

CostFunctionJ

List of values. Each element is a value of the chosen error function $J$ .

Example: J = ADD.get("CostFunctionJ")[:]

CostFunctionJb

List of values. Each element is a value of the error function $J^b$ , that is of the background difference part. If this part does not exist in the error function, its value is zero.

Example: Jb = ADD.get("CostFunctionJb")[:]

CostFunctionJo

List of values. Each element is a value of the error function $J^o$ , that is of the observation difference part.

Example: Jo = ADD.get("CostFunctionJo")[:]

CostFunctionJAtCurrentOptimum

List of values. Each element is a value of the error function $J$ . At each step, the value corresponds to the optimal state found from the beginning.

Example: JACO = ADD.get("CostFunctionJAtCurrentOptimum")[:]

CostFunctionJbAtCurrentOptimum

List of values. Each element is a value of the error function $J^b$ . At each step, the value corresponds to the optimal state found from the beginning. If this part does not exist in the error function, its value is zero.

Example: JbACO = ADD.get("CostFunctionJbAtCurrentOptimum")[:]

CostFunctionJoAtCurrentOptimum

List of values. Each element is a value of the error function $J^o$ , that is of the observation difference part. At each step, the value corresponds to the optimal state found from the beginning.

Example: JoACO = ADD.get("CostFunctionJoAtCurrentOptimum")[:]

CurrentIterationNumber

List of integers. Each element is the iteration index at the current step during the iterative algorithm procedure. There is one iteration index value per assimilation step corresponding to an observed state.

Example: cin = ADD.get("CurrentIterationNumber")[-1]

CurrentOptimum

List of vectors. Each element is the optimal state obtained at the usual step of the iterative algorithm procedure of the optimization algorithm. It is not necessarily the last state.

Example: xo = ADD.get("CurrentOptimum")[:]

CurrentState

List of vectors. Each element is a usual state vector used during the iterative algorithm procedure.

Example: xs = ADD.get("CurrentState")[:]

EnsembleOfSimulations

List of vectors or matrix. This key contains an ordered collection of physical state vectors or simulated state vectors $\mathbf{y}$ that may be observed. These are $H$ operator outputs, i.e. simulated observation states (called “snapshots” in reduced-base terminology). At each step index, there is 1 state per column if this list is in matrix form, or 1 state per element if it’s actually a list. Caution: the numbering of the support or points, on which or to which a state value is given in each vector, is implicitly that of the natural order of numbering of the state vector, from 0 to the “size minus 1” of this vector.

Example : {"EnsembleOfSimulations":[y1, y2, y3...]}

EnsembleOfStates

List of vectors or matrix. Each element is an ordered collection of physical or parameter state vectors $\mathbf{x}$ . These are $H$ operator entries, i.e. current states before observation. At each step index, there is 1 state per column if this list is in matrix form, or 1 state per element if it’s actually a list. Caution: the numbering of the support or points, on which or to which a state value is given in each vector, is implicitly that of the natural order of numbering of the state vector, from 0 to the “size minus 1” of this vector.

Example : {"EnsembleOfStates":[x1, x2, x3...]}

IndexOfOptimum

List of integers. Each element is the iteration index of the optimum obtained at the current step of the iterative algorithm procedure of the optimization algorithm. It is not necessarily the number of the last iteration.

Example: ioo = ADD.get("IndexOfOptimum")[-1]

Innovation

List of vectors. Each element is an innovation vector, which is in static the difference between the optimal and the background, and in dynamic the evolution increment.

Example: d = ADD.get("Innovation")[-1]

InnovationAtCurrentState

List of vectors. Each element is an innovation vector at current state before analysis.

Example: ds = ADD.get("InnovationAtCurrentState")[-1]

OMA

List of vectors. Each element is a vector of difference between the observation and the optimal state in the observation space.

Example: oma = ADD.get("OMA")[-1]

OMB

List of vectors. Each element is a vector of difference between the observation and the background state in the observation space.

Example: omb = ADD.get("OMB")[-1]

SimulatedObservationAtBackground

List of vectors. Each element is a vector of observation simulated by the observation operator from the background $\mathbf{x}^b$ . It is the forecast from the background, and it is sometimes called “Dry”.

Example: hxb = ADD.get("SimulatedObservationAtBackground")[-1]

SimulatedObservationAtCurrentOptimum

List of vectors. Each element is a vector of observation simulated from the optimal state obtained at the current step the optimization algorithm, that is, in the observation space.

Example: hxo = ADD.get("SimulatedObservationAtCurrentOptimum")[-1]

SimulatedObservationAtCurrentState

List of vectors. Each element is an observed vector simulated by the observation operator from the current state, that is, in the observation space.

Example: hxs = ADD.get("SimulatedObservationAtCurrentState")[-1]

SimulatedObservationAtOptimum

List of vectors. Each element is a vector of observation obtained by the observation operator from simulation on the analysis or optimal state $\mathbf{x}^a$ . It is the observed forecast from the analysis or the optimal state, and it is sometimes called “Forecast”.

Example: hxa = ADD.get("SimulatedObservationAtOptimum")[-1]

13.15.5. See also¶

References to other sections:

Bibliographical references: