0028271: Remarks to math_PSO usability - MantisBT

ID	Project	Category	View Status	Date Submitted	Last Update

0028271	Open CASCADE	OCCT:Modeling Data	public	2016-12-23 11:53	2023-08-01 15:06

Reporter	~~ssv~~	Assigned To	~~msv~~
Priority	normal	Severity	minor
Status	new	Resolution	open
Target Version	Unscheduled

Summary	0028271: Remarks to math_PSO usability
Description	This is to share my experience with PSO tool from math. 1a) It is not clear how to distribute particles. I was expecting that particles follow uniform distribution by default in the search space, but this does not happen. The client code is bothered with the necessity of such distribution. I understand that in some cases (like extrema) the distribution of particles is dictated by the specifics of an objective function. But for a general purpose we should follow [1] (since we refer to this work in the comments). 1b) Even if the user is forced to distribute particles manually, just assigning positions to them is not enough. It is also required to initialize BestPosition which is a bit weird (see code snippet below). 1c) If we had a default distribution strategy for particles, then, I guess, we can have a default instance of math_PSOParticlesPool structure not to bother the client code with creating it. To summarize 1a), 1b), and 1c), we lack default initialization strategy for particles, don't we? 2) Each particle stores the best function value in "Distance" field. "Distance" keyword cannot be used since PSO is abstract, and the extrema is just one of its application. "Fitness" and "BestFitness" are better terms. 3) I failed to understand the meaning of the following arguments in PSO: - const math_Vector& theSteps (ctor). - Standard_Real& theValue (Perform() method). It required digging into the code to clarify what is "theValue" and "theSteps". Probably, "theSteps" can be made optional? Notice that if "theSteps" is used for termination, then there is some strange calibration in the code "aTerminationVelocity = mySteps / 2048.0" which looks pretty unclear. 4) Why PSO_Particle uses double* and not math_Vector? It seems to be inconsistent with the signature of PSO. 5) Why the weights of social and cognitive determinants are hard-coded? My experiments show that with such values PSO may diverge in terms of particle locations (not all particles sit in the same minimum). I wanted to adjust these weights, but it is not possible. Example (showing how tricky it is to run minimization for y = f(x)): // Number of particles const int numParticles = 30; // Inputs math_Vector xMin(1, 1), xMax(1, 1); xMin(1) = 0.0; xMax(1) = 100.0; math_Vector xStep(1, 1), xDelta(1, 1); xStep = (xMax - xMin)0.01; xDelta = (xMax - xMin)/numParticles; // Structure which stores and manages particles math_PSOParticlesPool psoPool(numParticles, 1); // Choose initial positions for particles for ( int i = 0; i < numParticles; ++i ) { math_Vector x = xMin + ixDelta; PSO_Particle* pParticle = psoPool.GetParticle(i + 1); pParticle->Position[0] = x(1); pParticle->BestPosition[0] = pParticle->Position[0]; } // Prepare objective function AverageDistFunc objFunc(...); // Outputs math_Vector xOut(1, 1); // Run optimization double value; math_PSO PSO(&objFunc, xMin, xMax, xStep); PSO.Perform(psoPool, numParticles, value, xOut); Refs: [1] Vaz, Vicente. A particle swarm pattern search method for bound constrained global optimization // Journal of Global Optimization. 2007
Tags	No tags attached.

Test case number

Date Modified	Username	Field	Change
2016-12-23 11:53	~~ssv~~	New Issue
2016-12-23 11:53	~~ssv~~	Assigned To	=> msv
2017-07-20 15:30	~~msv~~	Target Version	7.2.0 => 7.3.0
2017-12-05 17:01	~~msv~~	Target Version	7.3.0 => 7.4.0
2019-08-12 16:37	~~msv~~	Target Version	7.4.0 => 7.5.0
2020-09-14 22:56	~~msv~~	Target Version	7.5.0 => 7.6.0
2021-08-29 18:53	~~msv~~	Target Version	7.6.0 => 7.7.0
2022-10-24 10:43	szy	Target Version	7.7.0 => 7.8.0
2023-08-01 15:06	dpasukhi	Target Version	7.8.0 => Unscheduled