We are trying to breed some parameter configurations controlling the search of a deduction system. Some parameters are integers, some are reals, some are boolean, and the most complex one is a variable length list of different elements, where each of the elements has its own (smallish) sub-set of parameters. Since we have Python competence and Python is already used in the project, that looks like a good fit. I've found DEAP and PyEvolve as already existing frameworks for genetic algorithms. Does anybody have experience with these and can tell me about the strengths and weaknesses of the two (or any other appropriate) systems?
If it helps: In our application, determining the fitness of the individual is likely the most expensive part - it will certainly be minutes per generation, and if we are not careful and/or rich (to buy compute clusters), could be hours per individual. So time taken by the rest of the GA is probably not a major factor - think "several generations per day", not "several generations per second".
Plug-and-play hyper-heuristics may help you. It can be implemented in any programming languages. You should be able to use again the formulation for your aim.
I hope it helps.
Patricia
Conference Paper Plug-and-Play hyper-heuristics: an extended formulation
Conference Paper P Experimental results -Evolving Evolutionary Algorithms sol...
Pyevolve is a good framework...we can tweak it to our requirements. But in your case if you have any prior knowledge in pycuda...i would recommend you to implement everything in parallel computing programming paradigm....even if there is an improvement in the algorithm in 10-100 folds, you will be saving a day. I dont know if we can blend the Pyevolve with pycuda....it definitely sounds complex to do so. So why not exploit the GPU for your case.
I thought of doing it recently....if i could implement it. I will let you know.
I too would recommend DEAP. It is very flexible and intuitive. Development is active and the authors are very responsive to queries on the Deap users google group. I have not played with the distributed aspect yet but it does not seem overly complicated to implement.
Jeannie and Ioannis, it is interesting links. Have you got any research papers explaining which problem DEAP has solved. Please.
Thank you
Patricia
Just to add to the comments about DEAP. I can highly recommend it as it is very straight forward to use. Distributed computing is easily added to DEAP by using SCOOP (which basically allows you to distribute a map operation over muliple cores/ compute nodes via MPI). Keep in mind Amdah's law though when you look a parallelism though. In a project I have recently worked on (paper in draft stages) we found that using NSGA-II on large chromosomes became a huge bottleneck so parallel operation of fitness, crossover and mutation did not help us much.
Patricia:
If you scroll to the end of the page on the attached link, there is a list of relevant publications:
Jeannie
https://github.com/DEAP/deap
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python