Ultimatum X, on 27 October 2014 - 04:54 PM, said:
Instead of full on separate Elo, I think a small modifier that could be from the design team's assessment of that mechs tier or (probably easier) a small modifier to the weight class Elo per mech, as opposed to it's own independent score.
That's all, great podcast it was fun hearing you speak on the MM.
That approach would certainly be technically possible, yes. It would end up being very similar to the group size modifiers we have in place currently for group queue.
There are a couple issues though that are enough to give me pause.
Firstly, I don't have a controlled baseline to data-mine relative mech strengths from. Preferably I'd have a few tens of thousands of matches where 12 locusts faced 12 spiders, a few tens of thousands of games of 12 locusts vs 12 hunchbacks, a few tens of thousands of games of 12 locusts vs 12 timberwolves, on so on for every mech type..
With this data, I would have high confidence that I could calculate meaningful bias values that accurately reflect the various advantages of each mech type. Without this data, I worry that we'd end up generating inaccurate bias values, or even worse, faking the data by using best guesses.
Coupled to this, mech effectiveness can change on designs whim. The IS quirk pass is an excellent example of this. Every time a mech is altered by design, we would need to recalculate its' Elo bias, and to recalculate its' Elo bias I need lots of data!
The Elo algorithm itself elegantly handles hidden variables like this already. It's mostly a matter of convergence time for tracking that many individual skill variables. I'll definitely give it some more thought though, as it seems there should be a clever way to take advantage of the huge datasets we're generating.