Optimizer

From SUMOwiki
Jump to navigationJump to search

The SUMO Toolbox contains a whole hierarchy of optimizers that can be used by several components (for instance the OptimizeCriterion or the OptimizerModelBuilder).

A list of available optimizers is found in src/matlab/tools/Optimization/optimizers. Note that some optimizers need some external code available in the extension pack (or the Matlab Optimization toolbox).

A couple of example configurations are listed below. These tags can be used in supported components as stated above. When a particular option is not given a sane default will be used. Full information about the several options can be found in the optimizer source codee, implementation and corresponding papers.

CMAESOptimizer

Evolution Strategy with Covariance Matrix Adaptation (CMA-ES)

<Optimizer type="CMAESOptimizer">
	<Option key="maxFunEvals" value="Inf" />
	<Option key="stopFitness" value="10" />
	<Option key="maxIterations" value="1000" />
	<Option key="TolFun" value="10" />
	<Option key="sigma" value="-1" />
</Optimizer>

DifferentialEvolution

Differential Evolution.

<Optimizer type="DifferentialEvolution">
	<Option key="strategy" value="3" />
	<Option key="weight" value="0.85" />
	<Option key="crossover" value="1" />
	<Option key="populationSize" value="30" />
	<Option key="maxIterations" value="50" />
</Optimizer>

DirectOptimizer

The DIviding RECtangles (DIRECT) optimizer of Donald D. R. Jones.

<Optimizer type="DirectOptimizer">
	<Option key="eps" value="1e-4" />
	<Option key="maxevals" value="200" />
	<Option key="maxits" value="100" />
	<Option key="maxdeep" value="100" />
	<Option key="showits" value="false" />
	<Option key="penalty" value="25000" />
</Optimizer>

MatlabGA

Matlab Genetic Algorithm

<Optimizer type="MatlabGA">
	<Option key="Generations" value="1000" />
	<Option key="EliteCount" value="2" />
	<Option key="CrossoverFraction" value="0.8" />
	<Option key="PopulationSize" value="20" />
	<Option key="MigrationInterval" value="20" />
	<Option key="MigrationFraction" value="0.1" />
	<Option key="Vectorize" value="on" />
</Optimizer>

MatlabOptimizer

Wrapper around fmincon and fminunc functions of Matlab Optimization toolbox

<Optimizer type="MatlabOptimizer">
	<Option key="maxIterations" value="1000" />
	<Option key="maxFunEvals" value="100" />
	<Option key="largeScale" value="off" />
	<Option key="diagnostics" value="off" />
	<Option key="functionTolerance" value="1e-4" />
</Optimizer>

MatlabPatternSearch

Matlab Pattern search (patternsearch function of Matlab Direct Search toolbox).

<Optimizer type="MatlabPatternSearch">
	<Option key="maxIterations" value="500" />
	<Option key="maxFunEvals" value="100" />
	<Option key="timeLimit" value="Inf" />
	<Option key="meshTolerance" value="1e-4" />
	<Option key="functionTolerance" value="1e-4" />
	<Option key="maxMeshSize" value="Inf" />
	<Option key="cache" value="on" />
	<Option key="cacheTolerance" value="1e-6" />
</Optimizer>

MatlabSimAnnealing

Matlab Simulated Annealing.

<Optimizer type="MatlabSimAnnealing">
	<Option key="AnnealingFcn" value="@annealingfast" />
	<Option key="TemperatureFcn" value="@temperatureexp" />
	<Option key="AcceptanceFcn" value="@acceptancesa" />
	<Option key="TolFun" value="1.0000e-6" />
	<Option key="StallIterLimit" value="500" />
	<Option key="MaxFunEvals" value="1000" />
	<Option key="TimeLimit" value="Inf" />
	<Option key="MaxIter" value="Inf" />
	<Option key="ObjectiveLimit" value="-Inf" />
</Optimizer>

MatlabPSOt

Particle Swarm Optimization (needs PSOt toolbox).

<Optimizer type="PSOtOptimizer">
	<Option key="maxiters" value="200" />
	<Option key="popSize" value="24" />
	<Option key="acc_const1" value="2" />
	<Option key="acc_const2" value="2" />
	<Option key="initialInertiaWeight" value="0.9" />
	<Option key="finalInertiaWeight" value="0.4" />
	<Option key="epochInertia" value="150" />
	<Option key="minGlobalError" value="1e-25" />
	<Option key="gradientTermination" value="150" />
	<Option key="errorGoal" value="NaN" />
	<Option key="typePSO" value="0" />
	<Option key="seedPSO" value="0" />
	<Option key="mv" value="4" />
</Optimizer>

Add a custom optimizer

In order to implement a new optimization algorithm one should only create one new class in src/matlab/tools/Optimization/optimizers. The class should be derived from the base class Optimizer and MUST implement at least the following functions (for example implementations, please see the provided optimizers):

  • A constructor accepting one parameter (the configuration xml tag) OR a varargin
  • [this, x, fval] = optimize(this, arg )

Accepts a function handle to optimize OR a SUMO model that needs to be optimized. x and fval are the resulting optimum(s) and associated function value(s). In addition the following methods COULD be implemented if appropriate:

  • When implementing a population-based optimization algorithm: size = getPopulationSize(this)
  • When the algorithm supports constraints: this = setInputConstraints( this, con )

It is advised to copy a simple existing optimizer and work from there. For instance, DifferentialEvolution is a good starting point. The optimizer can be used like:

<Optimizer type="CLASSNAME">
	<Option key="option1" value="1" />
	<Option key="option2" value="2" />
	...
</Optimizer>