How does best in slot actually work, sims dont reflect askmrrobots suggestions

So, I was really excited when I first found Ask Mr. Robot. I find simming very confusing and don’t fully understand why one setup would sim higher than another. I always wished there was an addon or website that could just scan all my available gear and tell me what the best setup is. So when I found Mr. Robot, I was hopeful.

I know the basics—like for an Orc, you want 6% hit and then try to stack as much crit as possible without going over the crit cap. But when I sim and mess around with different gear setups using that logic, some setups will sim higher than others, even when they go against those basic rules— and I can’t seem to understand why that would be the case. For example, having more than 6% and less crit hit will sometimes sim higher. Or I’ll mess around with my gear in-game and find the setup that gives me the highest “Damage 250–331,” Attack Power, and crit and think i found the most optimal arrangement, but then it sims lower than other setups with worse raw stats.

That’s basically the same issue I have with Mr. Robot. I don’t understand why it suggests certain things. I select Fury, input my talents, and slide the toughness slider all the way to DPS. Sometimes when I go and sim the suggested setup from Mr Robot, it does show a DPS increase—but then I’ll try different gear combinations myself and find setups that sim even higher. So I’m confused why Mr. Robot didn’t recommend those.

It also makes suggestions that don’t seem to make sense. For example, I have Rivenspike and Bone Slicing Hatchet, and Mr. Robot tells me to replace Rivenspike with Serathil—even though Serathil sims much lower on WowSims/Guybrush. Or for my tank setup, when I move the DPS/Toughness slider all the way to Toughness, it tells me to put Rivenspike in my off-hand—even though it’s the slower weapon, which you typically wouldn’t want.

So overall, I’m just trying to better understand how to make gear decisions for my Warrior—what the basic principles are, so I don’t have to rely on simming every time. And also, how to use simulators and Mr. Robot more effectively. I find myself stressing so much every time i get a new piece of gear, wanting to optimize my set.

In general I would need to see specific examples to comment more on why the optimizer made a specific choice – you can provide us a snapshot by clicking the “help” link next to the big “Best in Slot” section header, and copy the generated 32-digit code into a post here.

With snapshots of specific cases we can talk a bit about your questions like “why in this case did it choose this weapon instead of that one”, etc.

As for how it works in general: we have a mathematical model for each classic spec that takes into account every talent, stat, etc., and calculates how hard each of your abilities will hit, which abilities you would use and how often, etc. It is a much faster approach than simulation that lets us examine hundreds of thousands or even millions of gear+enchant+gem combinations in a few seconds. We much prefer this approach to simulation these days (we also use it for retail WoW) because it lets us examine so many more combinations and return a result in much less time.

Why do our results not always match another simulation website? That’s an easy one – because they are independently developed models of the game. A lot of assumptions need to be made to make a simulator work, and same for a mathematical approach like ours. In general you should get very similar results between different models of the game, but it is rare that you will get identical results.

As for which model is more correct… usually it’s too hard to tell. The amount of tests you would have to do in-game is prohibitive. And furthermore, simulations usually simulate a very simple case like a “patchwerk” fight. Thus the model already diverges from the game, making its results not perfectly accurate for most in-game situations. That doesn’t mean the simulation results are useless. What it means is that you should not worry so much about small variations in the simulation results – sometimes it will work out better in the real game, sometimes worse… and it’s too hard to figure out which it will be.

There’s a bit of an art to determining when a variation in simulation or calculation results is “small” and can be waved away, and when it’s more likely an error in the model. With specific examples we can comment further.

Hello, thank you for your quick and detailed response. I’ll do my best to follow up and explain everything clearly.

Here is one snapshot: c29c3a6fac8b4eaa8e41a4bd48974236

That result came from setting the slider all the way to Damage, selecting only the raid buffs/consumes that Guybrush Simulator also has available, and disabling enchant changes—so I’m only seeing what my current best-available optimized setup would be, without needing to buy or re-enchant any gear.

If you take Mr. Robot’s suggested set and sim it using Guybrush’s simulator, then make a few small changes, you’ll find there are other variations that sim higher. So why didn’t Mr. Robot recommend those instead?

For example, I’ve always been told that when using Diamond Flask, you should drop Blackhand’s Breath and keep Hand of Justice. When I sim with Diamond Flask and HoJ instead, the results are better. Another example: if I keep Diamond Flask and Blackhand’s Breath as Robot suggests, but replace Heroism Shoulders with Truestrike Shoulders, that sim shows higher DPS too. And if I then equip my Heroism Gloves (enchanted with +7 Strength), the sim improves further. So why does Mr. Robot choose the setup it does, instead of these higher-simming options?

Here’s another snapshot: 914a4ad7937e443aafc97748126bcc58, where enchant changes are allowed. In this case, it recommends equipping Serathil in my main hand over Rivenspike. But when I sim the same gear set with Rivenspike instead, it performs better and every warrior I ask says you would not drop Rivenspike for Serathil. Also, if I follow the same changes from the previous snapshot—Truestrike Shoulders, and using HoJ—the sim results are again better.

SnapShot: ded815c7f7074f93ad973801447530a4
Another Example, and clarification I would like to know…I use my Fury DPS Spec to tank, and when i set the optimizer all the way to the toughness side… it tells me to put my Rivenspike in my off-hand. When I do not see how one way or the other would increase toughness and every warrior says you would not want the slower weapon in your off-hand. So what exactly does Robot take into consideration for tank sets specifically?
So I’m just trying to understand: how is Mr. Robot coming to these conclusions, when other setups sim higher?

So my first question is this: when you simulate on that website, how much of a difference are we talking? Could you post the simulated DPS for each setup in question? Your link to that site does not seem to have your actual setup saved on it.

I can ask our person who wrote the warrior code to look at your specific questions about weapons.

Regarding the weapon speeds for classic warriors… I had some lengthy discussions with people about that when we created the classic optimizer. I actually re-wrote the warrior calculations 3 times to refine the model as much as possible and make sure I didn’t have errors.

I would have to dig into the simulator you are using to verify some things (and honestly I probably won’t do that, but if someone else does I’d be happy to discuss). The biggest thing I would want to know is if the simulator is giving warriors any rage from taking damage, or only using rage the warrior can generate themselves. My calculations only use rage the warrior generates themselves. I think this would potentially affect where you put fast/slow weapons in a dual wield setup.

Here are some things I said back when this was all fresh in my brain:

We’re doing everything based on single target damage, so whirlwind damage isn’t going to factor in very strongly - at least not enough to warrant putting a slower weapon in the main hand. Faster weapons, by any calculation I can come up with, just give more DPS for fury warriors who are using bloodthirst, even factoring in all the weird stuff going on like heroic strike queuing.

If there is rage left over, it uses ww. All my tests show using rage on ww for single target reduces dps.

Having a slow weapon in the main hand isn’t necessarily better for heroic strike damage. Each heroic strike does more damage, but the rage you give up in addition to the rage cost is higher.

If you have so much rage that you can replace almost every main hand auto with heroic strike, a slow weapon makes sense. You have to be taking damage for that to happen. Right now I’m not assuming the rage you get from taking damage, I’m optimizing based on the rage you can generate.

Bloodthirst, execute, and auto attacks make up so much of your damage that the actual weapon speed becomes a relatively minor point in the optimization.

And, here is a thread with some classic warrior discussion that would be worth a read:

Let me know if any of that makes you feel better about our calculations! Once you hook up an optimization algorithm to an adaptive mathematical model, you don’t always get the assumed result. I think sometimes people forget that simulators also have to make a base set of assumptions about what is happening in the game - and any difference in those assumptions compared to our model will push the results slightly one way or another. For warriors in classic, my conclusion after much discussion and experimentation is that the slow vs fast weapon question is actually a very small difference in damage, so it is easy for the results to go one way or the other based on a number of factors.

It is deff a substantial DPS difference… for example here is the robot suggestion snapshot
253ce5522980493fa475e95f0e91f43c that says to use serathil over rivenspike and put BSH in main hard which sims like 1200dps. Meanwhile Rivenspike MH and BSH OF sims ~1230dps.

Here is the screenshots of the sims…https://imgur.com/a/apY6Rgs

Like i said its just one example…

My robot tends to get maybe 90-95% of your gear optimized and is a good starting point, but like i said when i would then take there suggested set up and put it into a sim and then maybe change 1-2 pieces around to what i think would be better… i would get much better sims. So It makes me wonder why robot didnt suggest those in the first place.

I cant really answer your question regarding what the simulators do behind the scenes… all i know is i use the most popular warrior sim that everyone uses and recommends. The two are guybrush spreadsheet and wowsims. These are recommended in the popular Warrior Discord “Fight Club”.

You can see my previous reply to yellowfive above, there is a pretty big dps difference. Lets say it was just about speed, how come it tells me to replace rivenspike for serathil?

Also I would love more insight into the robots tanking suggestions. I tank in my fury dps spec, so would i just use the same profile and what put it all the way to toughness? What are the stat priorities used? Does it no consider dmg/threat at all if it is all the way on toughness? etc

1200 DPS to 1230 DPS is a 2.5% theoretical increase in DPS. You could argue about whether or not that is “substantial” as 2.5% is not a big enough difference to empirically verify in-game. You’d have to do the exact same fight a few hundred times with each set of gear with the same conditions. Once you add in the variability of real fights, you’d probably have to do at least 1000 fights with each set of gear.

Lets say the simulator over-estimates the damage you can do by 1% on average and our optimizer under-estimates the damage you can do by 1% on average. Now there is almost no difference between the effectiveness of the two models.

I am pointing this out to get across the limits of theorycraft. Both our model and the popular simulators/spreadsheets get very similar results. They only differ in some relatively minor ways. If, as I mentioned above, the simulator is assuming some rage from getting hit, that would certainly account for the difference. Without doing a full audit of the simulator’s code, I couldn’t say for sure why there is this relatively minor difference. I CAN tell you that I went over our mathematical model in extreme detail for many hours. And, I also have 15+ years of experience doing theorycraft for WoW. I haven’t spent much (any) time promoting our website in recent years, so it is expected that my model is less popular. I am arguing that doesn’t necessarily mean it is less accurate.

Determining which weapon should go in which hand based on speed is not an obvious relationship due to the factors discussed above. The only way I could find for me to reliably settle on calculations that always put slow weapons in the main hand is to give the player more rage than they can generate on their own. This might be the case at some points in a fight, but not all the time. I think it is sound to gear for the scenario where you have to rely on the rage you generate yourself.

My robot tends to get maybe 90-95% of your gear optimized and is a good starting point, but like i said when i would then take there suggested set up and put it into a sim and then maybe change 1-2 pieces around to what i think would be better… i would get much better sims

If you use my mathematical model to pick gear, and then use a different mathematical model to test it, this could certainly be the case. I would once again argue that it is good to have multiple independently developed theoretical models. With your examples here, I’d say that AMR is getting you to 97.5% of what the other model thinks is optimal. That is actually nearly in agreement for two models of the game developed totally independently.

As far as tanking, in that case we do assume you get a lot of rage from getting hit. Sliding the gearing strategy over towards toughness will favor defensive stats. We don’t use stat priorities, we estimate an amount of damage coming in based on how hard some of the bosses hit in classic. We then try to get you a setup that reduces that damage as much as possible. It is more like an effective hit point calculation at it’s core (with some added complexity), which we have found over the years to be the most effective for staying alive. If you put the slider 100% to toughness, we don’t consider damage/threat for the most part - we never completely ignore the damage/threat you do though, because that leads to some weird results.