Understanding and Pushing the Limits of the Elo Rating Algorithm

[Michel]

https://arxiv.org/abs/1910.06081

Nothing that wasn't already known unfortunately.

[Daniel Shawul]

How did they miss our work Paired Comparisons with Ties ?
We compared three draw models and found out Davidson fits chess data best -- which they reframed as their own K-elo if I am not mistaken.
Maybe I should alert the authors to our paper and makes them spell out what their unique contribution.

[Robert Pope]

How did they miss our work Paired Comparisons with Ties ?
We compared three draw models and found out Davidson fits chess data best -- which they reframed as their own K-elo if I am not mistaken.
Maybe I should alert the authors to our paper and makes them spell out what their unique contribution.

"The requested page could not be found."

[Daniel Shawul]

Let me try again: Paired comparison with ties

[Michel]

Let me try again: Paired comparison with ties

You should probably contact them so that they can at least refer to your work.

But I think there is essentially nothing in the paper. As you say k-elo seems to Davidson.

The only thing that could be considered slightly original is that they identify the draw model corresponding to the pseudo likelihood approach to the elo model with draws counting as the average of a win and a loss (i.e. the contribution of a draw to the log (pseudo) likelihood is counted as (1/2)log(w(1-w)) where w is a function of the elo difference). But this is a triviality.

[towforce]

Just my opinion, but it seems to me that:

1. optimal chess is a drawn game

2. therefore, at a certain level of skill, all games will be drawn

3. therefore, any rating system that doesn't allow for this is fundamentally flawed

[Fulvio]

https://arxiv.org/abs/1910.06081

Nothing that wasn't already known unfortunately.

tl;dr

Based on the ELO difference, the current algorithm gives the following probability of winning:
+0 --> 50%
+100 --> 64%
+200 --> 76%
+300 --> 85%
+400 --> 91%

What are values of the improved algorithm?

[Dann Corbit]

A while ago, Kaggle had a contest for improved game outcome prediction algorithms.
Uri did pretty well in the contest, if I remember correctly.