How do GA plugins iterate through individuals for evaluation?

[sempocoaguilar1]

Hi. As part of my PhD research, I am using agent-learning in python with grasshopper. By using ladybug as an evaluator, the performance outputs of it should then become the input for an agent-learning algorithm. In this sense what I am trying to do is very similar to what GA plugins like octopus or Galapagos already do, but instead of GA use AL.
Now my question is quite simple, How GA plugins iterate through various candidates and their performances? This implies quite a complex sequence. First, they have to manipulate the sliders determining the generative parameters for an individual, and then after a fenotype is generated, they have to wait for the evaluator to finish its own internal process and give a number representing the performance of the individual as output. How do the GA plugins do this coordination with the evaluator plugin?
I have tried these solutions to iterate through candidate geometries:

• https://www.grasshopper3d.com/forum/topics/iteration-in-python
• https://www.grasshopper3d.com/forum/topics/controlling-sequential-flow-with-time-module-python?commentId=2985220%3AComment%3A1644677
  Nevertheless there is the issue of efficiency. The algorithm has to wait for a given time (likely to be longer that the processing time for the evaluation) until it can explore the next candidate.

[Dani_Abalde]

@sempocoaguilar1

How GA plugins iterate through various candidates and their performances?

They have specific classes to represent individuals and populations, the sliders are converted to these, they are not directly manipulated within the optimization process. They iterate as everything is iterated, with loop statements. It’s not so complex, you create a population of individuals where each individual is a composition of the input values (genes), to each one you evaluate their fitness, you take the best ones, you mix them and mutate them, and you repeat with the new generation of population, until it converges or you stop it and then you return the meaning of those individuals/gene sets to the sliders or whatever.

The fitness function or performance evaluation is the fundamental thing in these algorithms, and it is what is defined in the gh definition by the user. The algorithm only searches the input space in the direction of maximizing or minimizing the result of this function.

[DavidRutten]

These previous posts may be helpful.

GH_Document.NewSolution crashing Rhino

window code.gh (8.4 KB) The attached file has a C# component which shows a window with one button. The button click event handler starts a loop with 100 iterations. Every iteration adjusts all sliders plugged into the component and then runs a new solution.

Dynamically updating slider

RandomOptimiser.gh (20.7 KB) [solver] This is a lot easier if you can just have a straight up loop, but that requires your code callstack originates from outside the solution process. So it pretty much requires a UI event to kick things off which in turn requires some UI. The script I wrote uses an _iterations counter which is decremented every time a solution runs. When this counter hits zero, we’re done processing.

[sempocoaguilar1]

Ok, but how do I do to iterate through BREPs for evaluation? Evaluators such as ladybug need the geometry output to be able to do their work. When I loop through a list of BREPs using python, it does so internally, so the grasshopper component only outputs the last BREP in the loop. Maybe I did not adequately express myself. More than how to control the sliders, my question is on how to control the visualization of the outputs (BREPs) so there is a coordination between the generator of geometries and their evaluator.