MATLAB GA reject individual after fitness function evaluation

Hello everybody,

I’m currently using MATLAB ga optimizer applied to an external software (Aspen Hysys). MATLAB generates the input (genetic algorithm individual), and then passes these values to the Hysys simulation. Hysys is used as a “hidden engine”, it calculates the output and returns as result the fitness value of the individual. Now, I would like to be able to reject an individual in case some of its parameters (computed in Hysys) don’t respect my indications. At the moment I simply set individual’s fitness function to 0 if this happens, but I would like this individual not to be stored in the actual generation, to avoid the risk of a large number of individuals with zero as fitness value.

Can anyone help me?

Answer 1

[Star Strider]

‘Now, I would like to be able to reject an individual in case some of its parameters (computed in Hysys) don’t respect my indications.’

I have no idea what that means.

The ga function allows linear and nonlinear constraints on the parameters, and parameter bounds. See for example: Optimize with Linear Constraints and Bounds and similar discussions in the ga documentation.

[Enrico Gallo]

since i am not optimizing a mathematical function, but a chemical plant on an external simulator (Hysys), i cannot use "a priori" constraints to avoid some conditions to happen, but i have to verify them after the external simulator has given a result. At this point, if the verification has a negative result (e.g. a compressor works in a incorrect way) i would like to reject this individual and try with another one. In this moment i'm not able to do so, but i simply assign the fitness value 0 to this individual. I'm worried to obtain a lot of individual with fitness value 0, invalidating the work of the ga algorithm.

I hope to have been more clear

[Star Strider]

You are optimising a mathematical model of a chemical plant. (It would be best to model it directly in MATLAB.) I have no idea what Hysys produces (a vector?) so I would write the fitness function to check the result of the simulation using specific bounds on the outputs, or more appropriately perhaps multiply them (dot function) by some sort of weighting vector, so that any unacceptable output is assigned a very high scalar value instead of 0, and return that to the ga function. (Perhaps multiply the compressor value by an appropriate-magnitude negative constant, so that a negative number for it would produce a high value and a positive number would produce a low value.) Since the ga function searches for a minimun fitness value, the individuals producing high values would not be propagated in the population.

[Enrico Gallo]

I cannot model the plant directly on Matlab, only a specific program as Hysys can model the complexity of such a system (that's the reason why it is used, even though it is extremely expensive). Hysys receives from Matlab a vector and returns a scalar number (the production output). Your idea is the same that i'm already using: Matlab import the negative value of the volumetric flow of production in order to minimize it (this is like maximise the absolute value of production). If the actual simulation doesn't satisfy my constraints, the output is automatically set to 0, and theorically those individuals don't propagate. My worries are related on the number of these individuals: if at the first generation more than 70-80% of the population present a 0 value, the optimization process won't work well (i'm actually worried by having less than 5% valid solutions as extreme case). I would prefer the possibility to refuse an individual and recalculate another one, until it satisfies all constraints (that, i repeat, have to be calculated after the simulation is completed). This process would be longer in terms of time, of course, but more robust. But i don't know how to do this.

[Star Strider]

I suggested setting it to a very high number rather than 0. Individuals with a very high fitness will be selected against, and those with a very low fitness will be selected for. That is how ga works.