jsboige
commented
3 years ago I can see the continuous control build failed because of c# language features. I guess this is because I'm using a more recent version of VS community. As I mentioned, I won't have time in the coming days to get back to it, but I suppose this is minor adjustments.
In the mean time, I think it would be a good idea to upgrade a few versions in the project, which would probably yield that build to pass. Most notably, I played a bit with your Benchmark.Net project, and the latest Benchmark.Net version has breaking changes with your code, but I'd be willing to do the updates.
giacomelli
ktnr
This is a very large pull request that introduces a new powerful feature to GeneticSharp: Metaheuristics.
It comes with numerous fixes, tweaks and improvements, which are not necessarily related, so in order to make it easier to understand, I will also make 2 other smaller pull-requests that this one contains, related to issue #82. Hopefully we can validate those preparatory intermediate pull-requests before this one is entirely validated. Those pull-requests are #89 and #88.
Note that there isn't much code coverage yet, comments are lacking too, and there is probably some more work to do before it can be fully integrated in master branch, thus my proposal to merge it in develop first. However, I believe it is clean enough for an initial review, and there are important heavy unit tests passing that should clear the way. This is also that I will have to postpone working on it for about a week, and I figured now that the upgrade has started yielding very promising results, you might be willing to give it a try a give me your feedback.
Now about it: Metaheuristics are a new kind meta-operators that will allow controling the evolution process through other native operators dynamically. In the last 20 years, a lot of succesful ones were introduced, often based on natural metaphors. They manage to better guide evolution through finely tuned exploration/exploitation tradeoff, and prevent early collapse to a local optimum, which standard GAs and GeneticSharp exhibits (see issue #47 for a discussion about it).
In order to integrate known Metaheuristics and corresponding capabilities, I believe the mealpy repository is going to be a reference, since it implements pretty much all of them.
For now, I have set to myself the objective to have one well known good Metaheuristic working, namely Whale Optimisation Algorithm (WOA), for which I strictly followed the easier original Matlab version. The initial result can be seen in a factory class where I intend to implement add other famous Metaheuristics when initial integration is cleared. In order to make using Metaheuristcs easier, I have started providing the corresponding fluent API that is demonstrated through that initial example, and will be used throughout the other factory examples.
Another distinct Metaheuristic, which I also implemented and tested is the Eukaryote MetaHeuristic. It is of a different kind, because it leverages the available dynamical operator mechanism at the chromosome level rather than the evolution operators themselves as with other Metaheuristics. My work on that was largely inspired by the following publication introducing geometric crossover for Sudokus. The corresponding metaheuristics allows spliting a regular 81 cells chromosome into a 9 rows subchromosome karyotypes, and then using on each of them ordered crossovers and mutations preserving permutations. This is illustrated both in the Unit tests and in the updated gtk# application controller, though on this one you have to select the crossovers and mutation manually using the Gui. The new Eurkaryote-based chromosome outperforms the existing Cells chromosome and prevents collapse to suboptimal solutions in certain cases, which was the intended behaviour. Note that it is still inferior to the recently improved existing PermutationChromosome, which has a reduced search space and doesn't need the chromosome splitting overhead.
Those two initial test-cases guided me designing all the capabilities needed for their implementation, and should provide robust examples on how to leverage the new capabilities: there is now a vast range of options available through the combination of Metaheuristics building blocks.
There is a lot of code-coverage to add now, together probably with a couple bugs to fix, and I will be willing to do most of the clean-up, though of course if you appreciate the contribution, I don't mind doing that work collaboratively.
About modifications to the existing code-base, I tried to keep the footprint relatively light. I encourage you to compare benchmarks with other branches. There shouldn't be too much impact. That is, of course, when using the default Pass-through Metaheuristic. When using a relatively heavy Metaheuristc with sub-components like the WOA, the overhead is more severe (more than 2X), but it is largely compensated by the fact that the search reaches much better fitness that the original GA wasn't capable of reaching with the suboptimal modal collapse aforementioned, and it does so with populations an order of magnitude smaller (I encourage you to analyse the initial compund unit test: the Metaheuristic based vastly outperforms the regular GA). Now it doesn't mean there isn't room for improvements . I believe the parameter passing mechanism, though functional, might be interesting to refactor both for speed and fluent usage.
Some of the other technical discussion will probably concern the other 2 pull-requests, which are preliminaries to this one, so I'll leave it at here now, with that already long description.
I hope this doesn't look too frightening, that you'll enjoy the contribution, and I'm definitely up for discussing what needs discussed.
Cheers