Metrics to characterize control performance without simulation

Hello,

I am trying to find a metric to tune the parameters of a variable structure controller but right now I am using metric computed after simulation and it is very long plus the metrics cannot be computed in some cases. Is there a metric that could be used to tune the controller without simulating the model and that could be computed in any case?

Thank you

[Mathieu NOE]

hello , do you have a simple working code we can use ?

IMO,performance is related to error signals and I don't see how you could have a somewhat equivalent info without simulating your closed loop on a given test signal

[Lucas]

I am trying to use PSO to optimize the parameters with a code like this :

% x0 : initial parameters
% x_best : optimized parameters
mdl = 'sys_2';
load_system(mdl);
x0 = %depends on controller structure
nvars = numel(x0);
lb = %depends on controller structure
ub = %depends on controller structure
n_iter = 100;
swarm_size = min(20, 5*nvars);
options = optimoptions('particleswarm', 'MaxIterations', n_iter, 'SwarmSize', swarm_size, 'InitialPoints', x0);
x_best = particleswarm(@(x)fitness_function(x, mdl), nvars, lb, ub, options);

Right now the fitness function is derived from simulation. In the future, I want to create an evolutionnary algorithm that optimize the structure of the controller and where this process would be an iteration of the algorithm. This is what I would like to find a faster way to optimize the parameters of the controller.

Answer 1

[Naga]

Tuning controller parameters without running simulations can be tricky, but there are ways to speed up or avoid simulations entirely:

1. Use simplified models, replace full simulations with analytical models or reduced-order approximations.
2. Evaluate controller parameters directly using stability margins, control effort, or robustness criteria.
3. Surrogate model, train a machine learning model (e.g., Gaussian Process) to approximate performance metrics based on a few simulations, then use it for optimization.
4. Add constraints or rules to exclude obviously unstable or infeasible parameters before simulation.

For your PSO code, you could integrate a surrogate model. This might look like replacing your current fitness function with a trained model or pre-checking parameter stability to avoid wasting time on bad candidates.