Convergence - Writing your own Genetic Algorithm Part 5 

Can't believe how few subscribers you have, this series has been amazing, keep up the good work!

Hello, Firstly I wanted to say that this is a nice and very clear video, I especially like that you blend theory and programming tutorial. I am personally not a fan of checking the improvement on the objective function as a way to create a stop condition as, in my experience, GAs tend to make huge improvements by doing rare jumps, in addition to the gradual stream of small improvements you illustrated, at least on some problems. I did some research on GAs for uni and I found that most GAs (the ones that allow fitness to decrease) can be seen as Monte Carlo Markov Chains, that are processes that don't converge but are very useful nonetheless, especially in statistics. I like to think that we should stop GAs the same ways we stop a MCMC process, by checking that the stationary distribution is reached (convergence in probabilities). In simpler terms, this notion corresponds to a sufficiently explored search space, unfortunately, methods to automatically check that the stationary distribution is reached in a general case is still a problem in computer science research. For now, it seems to me that we have to use properties of the problem at hand to create good stop conditions. Then, a local search method like gradient descent with a traditional convergence stop condition like the one you taught can be used to improve a bit the best solution(s) found. In Mouret and Clune 2015, they showed that MAP-Elite, a kind of GA that focus on diversity of solutions instead of on fitness optimization tend to produce even better solutions that traditional GA search, a finding that seemed to be confirmed by my own experiments. This suggests that we should not focus on fitness improvements alone, indeed...

Started watching the series tonight, and while I am anxiously awaiting more... I think you should now change the colour of GA in your name.

Very cool series. Just wanted to add that you should probably not compare floating points by equality, but check if the absolute difference is less than some threshold. Thank you for the clear explanations !

I usually use the number of generations or computing time to stop GA. Sometimes I use the convergence to restart search process of my GA, which I called adaptive restart GA. By the way, very good tutorial, clear and informative. Many thanks

Thank you for this wonderful video. However, I am having issue with references. I try to add them by clicking project, add references but nothing happens. even System one is not found. Is it caused by visual studio version you have used to create the project? I was able to run all previous examples by setting project path to lib folder of the project with no issue but this one is blocking me.

Awesome video! I was curious, which extension for VS do you use for the code formatting (eg. arrow instead of =>)?

Thanks for wonderful video. Can we display the sequence of town may be in console? I want to see different sequences.