I have been a long-time fan and user of AMR, but I am a little curious about how it works. I originally used it in BFA when it ran a sim of your gear each time and made recommendations. But after I’ve read on the forum about how it simulates the classes then has a database to pull information from. Does AMR come out to being the better general quick recommendation? Or is it still more accurate? I’ve not found much information about the APLs or how AMR works past extremely old information from 5+ year-old forum posts. I’ve noticed that with running personal simulations through SIMC comparing AMR recommendations to comparing gear using only SIMC that AMR seems to get close but never has the top recommendation. Does this come to a difference in APL or something else? If you’d like I can create a snapshot but I’m more asking from the theory crafting standpoint.
I wanted to add to this. I much prefer AMR over simc as even things like the pawn ratings are much more useful on AMR, but I am curious why some players live by simc when they do not support many things that AMR supports.
Back in the day we used simulations to rank gear, but now we are using mathematical models instead. We made this switch at the beginning of Dragonflight and have been refining it ever since.
With Dragonflight, the number of talent combinations really exploded (in the millions for each spec), so it became impossible to run enough simulations in a reasonable amount of time and for a reasonable cost to create our simulation-based statistical models.
This new approach still calculates and estimates actual ability use counts, damage, buff uptimes, etc., but it is very fast compared to a simulation. We can give you optimization results in just a second or two that adapt to any combination of talents, settings, etc. It also has the advantage of being way faster for us to update when the game changes. In the past we had to let simulations crank for days every time the game was patched – now we can get updates out to users on the same day for most patches.
As with the simulator we used to use, our model is completely independently developed, so while you’ll see very similar results in most cases to e.g. simc, you won’t get exactly the same solutions. There are a ton of assumptions that go into a simulation or mathematical model of the game so it is unlikely that you’ll see two independent models match exactly.
Most of the time the results are close enough that it’s tough to say which would actually perform better in-game. It would take too long to do enough trials of a controlled-enough test to say for sure – there are a lot of variables in-game that no simulator or calculation will capture anyway (e.g. random boss events, the timing of how your particular team does the fight, and so on), so there is a hard ceiling on how accurate any model can be. This is why we tell people not to think in terms of One Absolute Best in Slot list… Best in Slot is more like an “archetype” describing a cloud of thousands (or even millions) of functionally similar solutions – a different gem here, swap these two similar items over there, etc.
If you see wildly different suggestions that seem suspect to you, definitely let us know. But small differences? Anyone being honest with you will let you know that they fall below the threshold where any model can definitively tell the difference.
We do plan to update our blog sometime over the next few months (it fell by the wayside as we got behind trying to keep up with classic versions of the game as well), and we plan to make a few posts about various theory-related topics for people into that kind of stuff.
Thank you for answering with detail, I have always loved amr as it is the only really easily adaptable solution to gear sims. But I do have a bit of experience with simc and could see some differences at times. But the only information I ever found was extremely old, I think bringing your blog with more detailed information will help bring AMR more credibility. A lot of theorycrafters I know often say they don’t trust it due to lack of information. And that’s the only reason I’ve ever been given. And I’ve always recommended AMR!
Yeah I’ve heard that criticism a lot over the years. I understand it to a point… but for about six years we did give people all the information that they needed to verify the accuracy of our model. We had all the guts of our simulator on the web, and you could even modify it and it would dynamically compile and run your modified version right on the web – very cool! We had top-notch simulator log output with filtering and searching that anyone could use to dig into exactly how it was working.
Didn’t really matter… people just went around saying it was a black box anyway. It was honestly quite frustrating because we put a lot of effort into making it more transparent to people who aren’t computer programmers.
So we’re working on other ways to demonstrate to people that, yes, after 14 years of doing gear optimization, we do in fact understand how World of Warcraft works!
Until we get some more content out there I like to tell people: it’s actually way easier to make a gear optimizer that works than to convince a ton of people to use one that doesn’t.