3 ms·
Agreed. ELO is a crude but simple statistical model, and its primary benefit is that it's easy to understand and reason about. I'd love for there to be a multi
by Scene_Cast2 2y ago
Agreed. ELO is a crude but simple statistical model, and its primary benefit is that it's easy to understand and reason about.
I'd love for there to be a multidimensional player skill vector (aka "embedding") that encapsulates more information as well as a more nonlinear model to go on top of it.
For example, in many games, a team usually needs a support player, but a single number isn't enough to capture that for the purposes of matchmaking.
- Spawek 2y agoAuthor here: Elo is commonly used and most gamers understand it. It's nice to be able to show each player their Elo. I love the idea to use per-player embedding and to add a nonlinear model on top. I think it may improve the model dramatically, but I'm afraid that the model may be overfitting strongly as we have only 35 games in the historical for now. Anyway - I think I'll try this out.
- Scene_Cast2 2y agoYeah, 35 games is hard to work with. My rule of thumb is to not go above 1 parameter per 10 data points in order to prevent overfitting and memorization, so I'm not sure how well it will work in your case.