User avatar
ZINDI
Scoring and submissions update
Platform · 27 Sep 2024, 12:58 · 6

Hello Zindians,

As we come to the end of this challenge, we would like to first say congratulations for all your hard work up until now. We are excited to see how your solutions improve in the final moments!

We recognise that this has been one of the more difficult and confusing challenges on the platform, and that the problem statement and evaluation metric are significantly different from our more typical machine learning challenges. As such, we've decided to disable submission selection for this challenge; your final score and rank will be based on your best performing submission out of all submissions made.

As scoring will be on all submissions made in the course of the competition, we request that you please keep track of any changes made while iterating throughout the competition, to ensure reproducibility of your solution and results.

Additionally, private leaderboard scores will not be made available until 1-2 days following the close of the competition. This is to allow the Zindi team time for rescoring and validation of all submissions/strategies.

Best of luck as you finalise your work, and we look forward to seeing everyone on the private leaderboard!

Discussion 6 answers

This is incredibly unfair 3 days from the contest end.

I can only speculate about the causes, but this should have happen either week one, or, if this late, postpone the contest end at least one week to give people time to both read your message and react to it, otherwise you are just penalizing people for not being active until almost literally the last day.

I'm also unsure about what problem does this solve. I can only assume that most of us went with the idea that the private leaderboard was going to be like the public one, with disqualification, and planned accordingly. This change screams to me to randomely change parameters and pray than one of our solution passes the cut. Is some code better because it can produce a low score solution maybe 1 out of 20 or 40 times, instead of letting people judiciously chose the 2 most robust solutions their code can produce?

Also, how was this change verified and tested, for I can still see and chose two submissions in my submission list? Is it intended that I still have the option to chose when only the best one will be taken into account? This makes little sense to me.

And finally, there's this: https://zindi.africa/competitions/smart-energy-supply-scheduling-for-green-telecom/discussions/22584 where I lost every submission from before 5 days ago because of a site bug, with maybe a better one in there that what I currently have.

Maybe I'm alone in this, but even if I may benefit from it, I don't think this was the right decision to take, to whatever issue you claim are adressing.

27 Sep 2024, 17:04
Upvotes 7
User avatar
Juliuss
Freelance

The more the submissions the higher the probabilty many feasible solution would pass through. I wish I knew this earlier!

27 Sep 2024, 20:50
Upvotes 4

I agree with Daxida. Changing the rules in the middle of a contest is bad. Changing it 3 days before the end even more so. People who made full use of submitting many solutions every day have a big advantage and can now get lucky. Rather than encouraging conservative solutions that optimize for the worst case now the objective becomes to cover many possibilities with a wide range of submissions.

Also consider that not every entry satisfies the 1 sec strategy calculation time limit. I expect people will initially ignore this time limit and then optimize it for the two selected candidates.

I do see the issue that the shake-up could be huge because many top solutions may get disqualified for being infeasible or for going over the time limit, but this is not the way to solve it in my opinion.

Ideally for testing time limits you'd submit source and it gets verified and you get feedback if it satisfied it. Instead of having submissions infeasible, maybe it would have been better to use a metric that punishes going over the limit really hard, but doesn't give you 0 points. In case many top solutions would get disqualified, they would now still be ranked.

I also think, if you want to go ahead with such changes you should at least extend the time for the contest by a generous amount to allow people to adapt.

28 Sep 2024, 09:34
Upvotes 2

I think it is good that organizers realized there was an issue with scoring and tried to find a way around it.

However, I second @daxida and @tfriedel, that implemented solution is unfair. Some issues with running old solutions, like 1sec calculation time or traceability of old submissions, were already stated. I want to raise awareness of two more potential issues:

- what if somebody's heuristic is end-aware? In a sense it optimizes for the selected period, that is for the public leaderboard for 4 days. Such an approach does not care about the next 3 days and all the submissions will not work for private. Selected two would be prepared for 7 days, as the rules said. But now that this has changed, simpler (possibly lower-scoring) approaches without that will have tens or hundreds of submissions, while this will be penalized.

- what if somebody for the public leaderboard disabled the last three days (for whatever reason, to calculate local scores in the same way, to cut time, there could be multiple)? The person would be at a large disadvantage because the rules changed!

Please consider either extending time or continuing with multiple submissions, but add e.g.: 10-20-30 for the last day and use just those.

29 Sep 2024, 09:17
Upvotes 1

Don't forget the final private score will only contribute 50%. We all know that some lucky aggressive strategies will go to the top. Given the new rule, the more submittions you have, the more chances you can win the lottery. But are those strategies generated within 1 second? Are they for general cases or just work for special extreme optimal cases?

The report and code will account for the other 50%. I think it's fair for 50% vs 50%. As long as your solutions follow the rules and designed for general purpose, you still have the chance to win even you are not first in the private board.

Good luck to everyone!

29 Sep 2024, 16:34
Upvotes 0

I tried to point out, that now the lottery might be unfair to many (described who could be affected) and to be considered for report scoring you have to be at the top. This was provided in one of the other threads:

"Yes, you are correct. The report on the top solutions in the leaderboard (based on the final private score) will need to be submitted after the competition has ended." - nicolapiovesan