|On this page…|
When you use sdo.optimize to optimize model parameters (design variables), you must provide a MATLAB® function as an input to sdo.optimize. This function, also called a cost function, must evaluate the cost and constraint values for the design variable values for an iteration. (The cost and constraint functions are collectively referred to as requirements.) sdo.optimize calls this function at every optimization iteration and use the function output to decide the optimization direction.
The cost function can also be used for global sensitivity analysis. You generate samples of the model parameters and evaluate the cost function for each sample using sdo.evaluate.
The cost function must have:
Input — params , a vector of the design variables (param.Continuous objects) to be optimized.
(Required) vals , a structure with one or more fields that specify the values of the cost and constraint violations.
(Optional) derivs, a structure with one or more fields that specify the values of the gradients of the cost and constraint violations.
You perform the following tasks within the function:
Extract the current design variable values from params.
If the simulated response is required for evaluating the requirements, then simulate the model using the current design variable values.
Evaluate the requirements.
Specify the requirement values as fields of vals.
To use a cost function with sdo.optimize, enter:
[param_opt,opt_info] = sdo.optimize(@myCostFunc,param)
Here, myCostFunc is the name of the MATLAB function and param is a vector of the design variables.
Similarly, to use a cost function with sdo.evaluate, enter:
[y,info] = sdo.evaluate(@myCostFunc,param)
The software provides you with the following convenience objects that can you can use in the cost function:
Use an sdo.SimulationTest object, also referred to as a simulator, to simulate a model. The simulator allows you to simulate the model using alternative inputs, model parameter and initial-state values, without modifying the model.
You configure the simulator to log the signals needed to evaluate requirements and use the sim method to simulate the model. Then, you extract the model response from the object and evaluate the requirements.
Use these requirements objects to specify time- and frequency-domain costs or constraints on the design variables.
You configure the properties of the object and then use the object's evalRequirement method to evaluate how closely the current design variables satisfy your design requirement.
Use an sdo.Experiment object, also referred to as simply an experiment, to specify the input/output data, model parameter and initial-state values for parameter estimation.
You update the design variable values associated with the experiment using the setEstimatedValues method. Then, you create a simulator, using the createSimulator method, to simulate the model using the updated model configuration.
The function must take as input a vector of model parameter objects (param.Continuous objects) and, optionally, initial-state objects (param.State objects). These objects represent the design variables of the optimization problem. You obtain these objects by using the sdo.getParameterFromModel and sdo.getStateFromModel commands.
To access a design variable value, use:
param_val = p(1).Value;
Here, p is a vector of param.Continuous objects and p(1) is either a model parameter or an initial-state object.
sdo.optimize requires that the cost function accept only one input argument, params. However, you might want to use additional inputs. For instance, you could make the model name an input argument and configure the function to be used for multiple models. To call sdo.optimize and use a function that accepts more than one input argument, you use an anonymous function. For example, suppose myCostFunc_mult_inputs is a cost function that takes param, arg1, and arg2 as inputs. Then, assuming that all input arguments are variables in the workspace, you enter:
myCostFunc = @(param) myCostFunc_mult_inputs(param,arg1,arg2); [param_opt,opt_info] = sdo.optimize(@myCostFunc,param);
Additional inputs can also help reduce code redundancy and computation cost, given that the function is called repeatedly by sdo.optimize during optimization. For instance, if you use a convenience object in your function, you can create it once, before calling sdo.optimize. Then, you can modify the convenience object's properties as required within the function for each iteration.
The core of the function is where you evaluate how well the current design variables satisfy the design requirements. You can use MATLAB functions to do so. You can also use the requirements objects that the Simulink® Design Optimization™ software provides. These objects enable you to specify requirements such as step-response characteristics, gain/phase margin bounds, Bode magnitude bounds, etc.
Parameter-only requirements — Extract the design variable values and compute the requirement values.
For example, you can minimize the cylinder cross-sectional area, a design variable, in a hydraulic cylinder. See Design Optimization to Meet a Custom Objective (Code).
Model response-based requirements — Simulate the model using the current design variable values, extract the model response, and compute the requirement values.
There are multiple ways to simulate the model, including:
Using an sdo.SimulationTest object. You update the model parameter values using the simulator's Parameters property. Then, you use the sim method to simulate the model and extract the logged signals from the simulator that are of interest. For an example, see Design Optimization to Meet a Custom Objective (Code).
In parameter estimation, you can use the createSimulator method of the sdo.Experiment object to create the simulator. Before creating the simulator, you update the experiment with the current design variable values using the setEstimatedValues method. For an example, see Estimate Model Parameters Per Experiment (Code)
Linear model-based requirements — Update the model with the current design variable values, linearize the model, and compute the requirement values.
Your function must return a structure containing the cost and constraint values for the current design variables. This structure must have one or more of the following fields, as required by your optimization problem:
F — Cost value.
Cleq, Ceq — Nonlinear constraint values. The solver satisfies Cleq ≤ 0 and Ceq = 0.
leq, eq — Linear constraint values. The solver satisfies leq ≤ 0 and eq = 0.
If you have multiple constraints of one type, concatenate the values into a vector, and specify this vector as the corresponding field value. For instance, if you have a hydraulic cylinder, you can specify nonlinear inequality constraints on the piston position (Cleq1) and cylinder pressure (Cleq2). In this case, specify the Cleq field of the output structure, vals, as:
vals.Cleq = [Cleq1; Cleq2];
For an example, see Design Optimization to Meet a Custom Objective (Code).
By default, the software computes the cost and constraint gradients using numeric perturbation. However, you can specify the gradients and return them as an additional output. This output must be a structure with one or more of the following fields, as required by your optimization problem:
F — Cost derivatives.
Cleq — Nonlinear inequality constraints derivatives.
Ceq — Nonlinear equality constraints derivatives.
You must also set the GradFcn property of the optimization option set to 'on'.
Simulink Design Optimization does not support multi-objective optimization. However, you can return the cost value (F) as a vector, representing the multiple objective values. Using this approach does not halt the optimization. Instead, the software sums the elements of the vector and minimizes this sum. The exception to this behavior is if you are using the nonlinear least squares (lsqnonlin) optimization method. The nonlinear least squares method, used for parameter estimation, requires that you return the error residuals as a vector. In this case, the software minimizes the sum square of this vector.
If you are tracking multiple signals and using lsqnonlin, then you must concatenate the error residuals for the different signals into one vector. Specify this vector as the F field value.
For an example of single objective optimization using the gradient descent method, see Design Optimization to Meet a Custom Objective (Code).
For an example of multiple objective optimization using the nonlinear least squares method, see Estimate Model Parameters Per Experiment (Code).