Plackett-Luce question

[kejadlen]

Hello there, I wanted to use this algorithm in Ruby and started porting the code over, but I have a question about your implementation of Plackett-Luce.

openskill.js/src/models/plackett-luce.js

Lines 15 to 26 in 87a3074

	const [omegaSum, deltaSum] = teamRatings
	.filter(([_qMu, _qSigmaSq, _qTeam, qRank]) => qRank <= iRank)
	.reduce(
	([omega, delta], [_], q) => {
	const quotient = iMuOverCe / sumQ[q]
	return [
	omega + (i === q ? 1 - quotient : -quotient) / a[q],
	delta + (quotient * (1 - quotient)) / a[q],
	]
	},
	[0, 0]
	)

There's a reduce after a filter, so wouldn't the q index be incorrect when indexing into sum_q and c, or am I misunderstanding the code here?

[vivekjoshy]

I had the same issue porting to python, but I ended up using for loops instead. Maybe this will give you more insight into what is going on in the code.

https://github.com/OpenDebates/openskill.py/blob/ad8dc00f0dc15e00a39d3a84840fd0f85ac85bfb/openskill/models/plackett_luce.py#L20-L38

[kejadlen]

Thanks - I'd actually been using your codebase as well for a reference for my Ruby port already!

Poked around a bit more, and I think it doesn't matter, since the teams are sorted by rank coming into the model already, so the filter is only removing teams from the back of the list and the indices are the same in the end.

          omega_sum, delta_sum = team_ratings
            .filter {|_,_,_,q_rank| q_rank <= rank }
            .map.with_index
            # .filter {|(_,_,_,q_rank),_| q_rank <= rank }
            .inject([0, 0]) {|(omega, delta), ((_,_,_,q_rank), q)|

[philihp]

Hi!

You have a keen eye to spot this!

Apologies for missing this and thank you @daegontaven for responding so quickly. You may also be interested in the source code in the zip file at https://www.csie.ntu.edu.tw/~cjlin/papers/online_ranking/ which I ran to generate a lot of my test cases, and then I wrote a lot of my code in a functional style, but verified that my outputs were the same.

It's not an issue with sumQ because both of them have the filter where q <= i but I guess I reversed the order on the util vs the model! Since you found it, would you like to submit a PR to flip one of them and make it ever so easier to read?

I think you've found a problem with a though. Usually this just returns an array of 1 for every team, because there's always 1 team at every rank. I think we might have issues with something like rank: [1,2,2,3,3,3]; a bit contrived, but that would explain why it has never before been an issue. I'm gonna reopen this so I don't forget to take a look sometime.

[philihp]

Opening this issue up to anyone on hacktoberfest — good start would be to create a test case with ranks [1,2,2,3,3,3] and identify an issue.