That IS the problem, if you’re adjusting your experiments using something you “think” you know, you’re getting confirmation bias and you could get trapped in a state with “numbers” that look good but real results that are terrible. There is a reason why randomized trials are a gold standard in determining efficacy of drugs.
1 Like
lol, their MM “algorithm” has zero learning involved. They just average the MMR. And then because these MM results (and game outcomes) are highly random (and the MMR increase/decrease is solely determined on whether your team wins or loses) what they get is large oscillations in their skill assessments making it hard to converge to anything.
Back when they had the resources they could actually build a decent performance-based matchmaking system using ML, but they never had the right people to do it and now they definitely don’t have the resources. Hell, look at the state of AI, it’s as if an 8-yr old wrote it and it’s been like that forever.
Actually i think AI at release was “okay”, but then they tinkered with it and made it worse.
1 Like
Yeah the new AI is worse IMO.
Back to the topic at hand. I think the matchmaking is worse but I will say it would probably work if they had the numbers. Pretty much they needed the people for this system to work.
1 Like
Right, they had an ok one since the release. Its after some major update a couple of years ago (or however long that’s been) where it just became atrocious.
1 Like