Bug: other islands do not increment their generations

[theahura]

I am observing that when I have multiple islands, only island 0 ever increments its generation count:

2025-10-12 15:24:02 | openevolve.database  | INFO     |  * Island 0: 5 programs, best=0.8743, avg=0.5215, diversity=1742.58, gen=6 (best: ee76dc80-4872-401b-a364-cb832b802afb)
2025-10-12 15:24:02,188 - INFO -  * Island 0: 5 programs, best=0.8743, avg=0.5215, diversity=1742.58, gen=6 (best: ee76dc80-4872-401b-a364-cb832b802afb)
2025-10-12 15:24:02 | openevolve.database  | INFO     |    Island 1: 2 programs, best=0.8743, avg=0.8709, diversity=806.50, gen=0 (best: 82d83ba0-de71-4b38-8800-b2f4a945372c)
2025-10-12 15:24:02,188 - INFO -    Island 1: 2 programs, best=0.8743, avg=0.8709, diversity=806.50, gen=0 (best: 82d83ba0-de71-4b38-8800-b2f4a945372c)
2025-10-12 15:24:02 | openevolve.database  | INFO     |    Island 2: 0 programs, best=0.0000, avg=0.0000, diversity=0.00, gen=0                                                                                                                                                                                                                                                           2025-10-12 15:24:02,188 - INFO -    Island 2: 0 programs, best=0.0000, avg=0.0000, diversity=0.00, gen=0
2025-10-12 15:24:02 | openevolve.database  | INFO     |    Island 3: 0 programs, best=0.0000, avg=0.0000, diversity=0.00, gen=0                                                                                                                                                                                                                                                           2025-10-12 15:24:02,188 - INFO -    Island 3: 0 programs, best=0.0000, avg=0.0000, diversity=0.00, gen=0
2025-10-12 15:24:02 | openevolve.database  | INFO     |    Island 4: 2 programs, best=0.8743, avg=0.8709, diversity=806.50, gen=0 (best: 90fce5eb-3989-4a5d-9750-85008bf75877)
2025-10-12 15:24:02,188 - INFO -    Island 4: 2 programs, best=0.8743, avg=0.8709, diversity=806.50, gen=0 (best: 90fce5eb-3989-4a5d-9750-85008bf75877)

Notice that Island 1 is on gen 6 and everything else on gen 0. I've seen this go pretty high, island 1 on gen 100+ and the rest on 0.

This seems like a bug, but I'm uncertain if it is a bug in the counting or a bug in the actual generation. I think https://github.com/codelion/openevolve/blob/main/openevolve/process_parallel.py#L675 is incorrect, because it will always prioritize island 0 and as long as island 0 returns first it will only ever iterate on that island;

and second, I think https://github.com/codelion/openevolve/blob/main/openevolve/process_parallel.py#L529 may be wrong, since it looks at the number of completed iterations instead of which island the completed iteration actually came from.

Note: claude code helped me with this analysis, so take it with a grain of salt.

[theahura]

Added some logs to see where iterations are being submitted. It looks like it is submitting to different islands, but regardless of which island is being submitted to only the island 0 generation count increases

[theahura]

I never see the 'Switched to island' log on line 522 in process_parallel.py 🤔

[theahura]

O i see the bug I think
https://github.com/codelion/openevolve/blob/main/openevolve/process_parallel.py#L429
programs_per_island = max(1, max_iterations // (self.config.database.num_islands * 10))
and

                    # Island management
                    if (
                        completed_iteration > start_iteration
                        and current_island_counter >= programs_per_island
                    ):
                        self.database.next_island()
                        current_island_counter = 0
                        logger.debug(f"Switched to island {self.database.current_island}")

I am running without a max_iterations set, which defaults to the very large 10k max_iterations. I have 5 islands. So programs_per_island is 10000 // 50 = 200. I think that means it has to run 200 iterations before any other island gets a chance?

I guess I'm also sorta confused why the islands don't seem to run in parallel. I see migrations between islands but the islands themselves do not actually seem to process 🤔 (I also recognize that I have a very poor understanding of how this system is parallelizing so all of the above could be wrong)

[theahura]

Hardcoding programs_per_island to 10 did 'fix' the issue, in that I am now seeing the generation count increase for other islands...unsure if this is actually a correct 'fix' though

[codelion]

The hardcoded programs_per_island can be made into a config it was added to make islands have some progress before initialization as without it there isn't too much diversity in the pool.

[theahura]

for now im just patching openevolve but would love if this was upstreamed as an actual configurable. I'd also love some info on what specifically programs_per_island is supposed to do, im not sure i understand why that was added / why that would impact diversity

[codelion]

You can send a PR for the fix I can take a look. I think it was originally added to avoid the situation where all islands end up with very similar programs. So, at the beginning we only have 1 initial program, we can copy that to all the islands and start the evolution at each island but since everything else is almost the same in the beginning it is likely they all evolve to similar distribution of island populations. It was added to do at least a few iterations on each island before initializing the new one so that there are a few different programs to choose from.

[theahura]

o interesting. How does that play with the migration parameter? I thought migration was mostly how things got initialized

[codelion]

migration should eventually help, I think this was meant to be used only during initialization of the island. once initialized it is okay if the populations go on their own with different iterations. I have only used islands with the config where number of islands is equal to number of parallel workers. In other cases, it may give unexpected results. We can fix it if we test more and see what config would make sense.

[theahura]

thanks for the quick review; would also love if this could get released through pypi soon!

[theahura]

wait @codelion which feature is used for setting parallel workers? is it the parallel_evaluations config setting, or something else?

[codelion]

it is parallel_evaluations config setting

new pypi released with version 0.2.18