Several factors merit their consideration:

1. Population size

2. Computer Hardware configuration

3. Complexity of fitness function.

4. No. of Decision variables and constraints.

One of the ways that I know is to have a moderate level of threshold. Because too low the value of threshold is, too many iterations and too much of time.

GA's take time especially if the individual takes 1708 iterations. What hardware are you running it on?

Check your hardware memory speed? I personally experienced slower iterations due to my hardware.

the speed of GA depends on with how much time the individual fitness function is solved. suggesting you to use parallel computing of Matlab if your hardware has 4 or more cores.

Time of execution is dependent on three things:

The algorithm here is GA, while your choice of implementation seems to include a set of nested loops in MATLAB. I don't suppose you have a wide range of choices for your hardware. So your aim should be improving your implementation.

Optimization is the key in here. You seem to be evaluating the fitness of each chromosome in the population by means of a for-loop. Obviously this calculation has no interdependency and hence could be parallelized. You could launch a separate process -- one for each chromosome -- for fitness evaluation (I find this operation pretty simple in Python). I would prefer multiprocessing to multithreading in this case.

Dear Sunita Parinam,

Basically, computational time of optimization algorithms is strongly depends on your computer hardware configuration and also the considered fitness function. However, I think the aforementioned computational time is unusual for least squares problems. Such computational time can belong to the problems with several million rows. To address the issue, execute your fitness function (input values can be randomly generated) and then check how did it take long. If it take long time so you can conclude that your fitness function is computationally expensive, or there is a problem in your coding.

Several factors merit their consideration:

1. Population size

2. Computer Hardware configuration

3. Complexity of fitness function.

4. No. of Decision variables and constraints.

One of the ways that I know is to have a moderate level of threshold. Because too low the value of threshold is, too many iterations and too much of time.

Thank you all for the replies

I am using Windows 7 home basic (32-bit) operating system, 4GB RAM, intel core i5 processor 650 3.20 ghz

depending upon the feedback i think i need to check my fitness function as i tried running it on other hardware and it is taking same time

Looks like your algo is getting stuck in local minima, try tuning the parameters mutation rate, thresholds bla bla..

Your population size is small, your hardware is adequate for a single computer, your fitness function and the number of decision variables are the remaining factors according to Jayaram.

@ Amol P Bhondekar...Thank You Sir 

I tried changing the threshold value. When i decreased it, my Merit function gives values in thousands and when i increase the value of threshold, it gives less value of merit function which is acceptable. But it is not having any effect on the program run time. 

@Graeme Smith... Thank You Sir

My no of decision variables are 28 which I have kept constant till now. Should i change the no of layers from 28, I think i should try to modify my fitness function and other parameters like mutation rate as mentioned by bhondekar Sir

Hi, how much time does calculating one solution's fitness take when it is the sole calculation?

By the way, isn't it wrong to have 20 solution candidates to decide 28 parameters? Did I misunderstand something?

@Ozan Özışık 

Sir one solution consists of 28 parameters (i.e. thicknesses) and my population size is 20 i.e. 20 solutions (each containing 28 variable values/thickness values)

My wavelength range for calculating transmittance consists of 61 wavelength points. So One Solution is checked for 61 times and gives 61 transmittance values one for each wavelength. Then next solution is checked again for 61 times and so on. One fitness value is calculated for each wavelength and finally summated to form one Merit function for one output transmittance (consisting of 61 wavelengths). This is repeated for every solution set. Time to calculate one merit function is very less of the order of few seconds but for one generation its taking a long time i.e. approx. 2 hrs for 10 iterations/ generations.

Few seconds is much time for one function calculation but it does not sum up to 2 hours for the whole population of size 20 and 10 generations. There might be some unexpected additional time because of selection, crossover, migration etc. procedures, this can be checked by using an easy-to-calculate merit function. You might change parameters in optimtool according to the results or find a way to optimize your real function. Are there any methods for approximate calculation of transmittance?

Thank You for the reply Sir. As per I have seen the literature, to calculate transmittance, we have to use characteristic matrix method. I will check more. 

When I am trying to calculate transmittance without using optimtool, It takes only a few seconds but with optimtool, it takes a long time. I think it is not converging properly due to inner loops only.  The problem is I am trying to use static arrays instead of loops but then the program is not giving the correct results. 

@Ariadne Tsambani...Thank You mam, I will check.

Dear All,

Earlier i was facing problem of my program taking long time (approx 2hrs) for 10 generations to complete but now it is solved after i modified my program. Now my program terminates in just few seconds and converges when average change in merit function value is min (I have taken default generations -100 now).

The current problem I am facing is although the value of merit function I am getting is minimum but the desired transmittance curve and obtained transmittance curve are not matching at all. 

I tried to check the value of merit function at each transmittance value whether it is closer to the desired or not. i have observed that the obtained transmittance value for each wavelength is not even closer to the desired in some cases. Therefore the curves are not matching. 

It means the either GA program is not calculating the thicknesses properly-one reason for the same can be - not giving the input thickness range proper. for which I have tried different thickness ranges-lambda/2, lambda/4, etc. 

Or the possibility is the merit function is not proper. 

In the both cases I am not understanding  what range I should give as I have tried many ranges e.g. lambda/4 or lambda/2 thicknesses. I tried to check the merit function at each step manually but when the transmittance values are not closer to the desired then should  I work on the thickness range. or It is the work of merit function to match the desired and obtained results. So should I concentrate on merit function? 

I have read a paper in which they are using a penalty function and objective function is calculated objective function + penalty function. Do I need to use the penalty function also? If yes, how to decide this? I am attaching the paper 

I am not understanding how to proceed further ans select the merit function. Please help is possible. 

Thank You.