That helps a lot, good looking out. I am a casual raider, trying to make my way to ‘know it all bad ass’ like yourself. The fact a lot of this stuff still makes limited sense hardly implies I am lazy. It simply means I am not yet a bad ass. I think I kind of got it though.
I couldn’t find where this may have been stated - but where is the data coming from? Is it from sims, logs posted to askmrrobot, or logs posted to warcraftlogs?
This is all based on simulations done with the Ask Mr. Robot simulator.
what is this bullshit weight that no one can understand, seriously, you guys have gone to shit
this is the last time I purchased a payed subscription to the website
I WANT TO SEE MY PROPER WEIGHT NOT SOME BULLSHIT LINES
kungur… this idea that stat weights are the “proper” way to rank gear is not true, it is just one way to rank gear, and not that great a way at that. We wrote an article about why we stopped using them back in december: http://blog.askmrrobot.com/machine-learning/
Also, we’ll be adding a button in the near future for you to export stat weights to e.g. Pawn, if you really like stat weights. It can be useful for a quick and rough estimate in-game right when an item drops, which you can then double check using the complete optimizer on our site.
If I understand right, the left and right edges of the bars represent the lowest and highest observed quantities of a stat as found in the “See Full Stat Details” graph; considering only those bars which surpass 99% or 96% of base-cast dps or nps respectively.
I have a few technical questions to try and understand this better, to look for opportunities to better refine it and also make it more… comfortable… to players who don’t understand the higher fidelity you are trying to achieve:
The full-details graph has just a few “steps” for quantities of each stat. For example, there could be 0, 1000, 2000, or 3000 of each stat; but never a random quantity in-between like 1234. I am curious if there is a specific reason for this, and if results could be better refined by either: (a) having smaller fixed increments; or (b) choosing truly random quantities of any integer value. I don’t expect either approach would drastically change the best-case upper bound, but it might influence the marginal 99/96% case and prevent a situation where several stats appear to have identical lower bounds.
Each full-details iteration must have some margin of error. What is it, and is it being considered when capping the 99/96% cutoffs for the main chart?
Suppose class XYZ normally has a high requirement for critical strike. In the simulation, one trial is found where at least 99% of best-case dps is achieved despite having very low critical strike. In all other trials, having such low critical strike results in suboptimal dps. Is it appropriate to draw the conclusion that very low dps is actually okay, or would there be a concern that this is just a random error that ought not to reflect on recommendations?
I’ll go through your questions in reverse order… from easiest to hardest to answer!
So your last point first – this doesn’t really happen in practice. The way that our statistical process and simulation works, it doesn’t really generate random/unexpected “outliers” – if a low-crit set of gear simulates to high DPS once, it will always simulate to high DPS (unless we change the simulator). Let me know if I missed part of your point, but I don’t think we need to worry about that.
The data used to render the stat display is not direct simulation results. It is generated from the predictive model that we create from a sampling of simulation data. This predictive model describes the n-dimensional stat “surface”, and always returns the same value for any specific combination of stats. While any predictive model created from a sampling of all possible gear combinations will have less accuracy than any particular “point” simulation of a specific setup, it has two properties that are desirable for both visualization and ranking gear. Firstly, it is very fast (near-instant). Secondly, it “smooths” the data a bit, which is highly desirable for display and ranking – this gets rid of “noise” that experimental data always has, preventing undesirable effects when ranking. For example, were you to use simulation results directly to rank two nearly identical sets of gear, it might “flip” them, even though one is clearly better because it has more of everything. Our predictive model will almost never do that – it uses statistics to describe the trend in a smooth way.
We use specific points when generating the stat graph so that we evenly cover the entire stat space of interest, and so that the number of data points stays sane. Even though we can quickly generate these graphs via the predictive model, it’s still not “free” – there are CPU and memory costs associated.
One idea I really want to get across better in the future: you don’t want to worry too much about the specific stat values. Saying that the best sets of gear tend to have between 9000 and 12000 haste rating is about as “specific” as you want to think about your stats. There are so many other variables that are going to move that around on you that trying to pin that upper/lower bound to more specific values is both misleading and of limited usefulness.
With so much of your gear’s value locked up in these non-stat effects now (legendaries, set bonuses, trinket effects, relics, enchant procs, etc.), you really need the full statistical treatment we give with best-in-bags and best-in-slot where we predict the interaction between your stats and special effects and adjust accordingly. The “stat problem” and the “special effect problem” cannot be solved independently anymore – they are all now the “gear problem.” A lot of people like to have “goals” for their stats… it just “feels” better to say “I want to go for X haste, OK I have that, now go for Y crit, …” You just can’t do it like that if you care about min/maxing.
I understand your first two answers (my last two questions) well.
Regarding the third answer (first question), consider the following scenarios:
Scenario A: The simulation uses increments of 1000 stats to conserve cpu power while covering the entire space. It finds that >99% dps is achievable with various stat combinations having 1000 or more haste, 1000 or more crit, and 1000 or more versatility. All three have differing upper bounds, but they all start at the same lower bound.
Scenario B: The simulation uses increments of 100 stats despite requiring more cpu power. It finds that >99% dps is achievable with various stat combinations having 800 or more haste, 1000 or more crit, and 1200 or more versatility. There is no change to upper bounds (vs scenario A), but now their lower bounds differ slightly on the main chart.
I’m curious if scenario B would have a noticeable impact on user perceptions. Obviously it doesn’t greatly change AMR’s recommendation; so from a rational calculation point of view the extra cpu effort isn’t worthwhile. But could this give users a more comfortable perception? It might be that those extra clock cycles make users feel like AMR is more correct despite you and I both knowing that higher fidelity has already been achieved from dropping stat weights in the first place.
It actually would have zero impact on the recommended gear – the data points generated for these stat displays are not the data points used to create the predictive model for ranking gear.
I’m not sure if making the stat display have higher fidelity and thus track what the optimizer is doing more closely would make a difference to people or not… hard to say without trying it.
So, the purpose of the stat goal chart is for us to see what gear / stat balance we should optimally be and provide us a general range that’s signified by the light colored bar. I’m going to have to agree with another poster saying that the dark bar confuses people. Also, yes, I can just mouse over the percentage to get the numeric value, but if I’m supposed to use that chart for reference, and there’s no visual indicator where I fall in that range on the chart, the format of information should be the same at first glance.
I guess what I’m frustrated with is that I am clicking “Best in Bags” so that I could get a simple answer. But instead, I find that I could “possibly” switch out X pieces because I fall into this large stat range and it’s about the same. Is there something that would help me see which could possibly be switched? Or if I equip this legendary or legendary pair, what gear would work best to keep those stats balanced? If it’s all manual, I’m not sure what the benefit of the premium would be. My favorite feature was Best in Bags for it’s simplicity.
The purpose is to give people some context for what Best in Bags is doing. Since gear optimization doesn’t really follow simple stat priorities anymore, BiB could move your stats around in non-intuitive ways.
You can still just do what BiB recommends without question and perform top-notch in-game. We were just getting many questions like: Crit is my best stat, but BiB just removed crit from my gear to get a different stat, why? The answer is because there is because simple stat priorities aren’t really a thing.
Having so much of your gear tied up with special effects makes the problem more complicated. What stats you have ends up being a side-effect a lot of the time.
If were were to put a marker on that graph to show where your current gear is… and then BiB picks a set that moves you into a stat range that it shows as probably sub-optimal, people would be even more confused! Even though that might be a totally valid optimization.
Look at it as a bar chart with an error bar at the end. That is how it is intended to be read, just to give people a rough idea of the relative value of stats.
I too would like to see where my current stats lie on the graph bar.
@yellowfive please, make this picture to show where you are at the moment… not only “what is good” for you.
I’m a bit conflicted on how I feel about the way the data on stats is presented.
I like the fact that you don’t give too much weight to weights (eheh), because they serve no more than a linear predictor in a neighborhood.
At the same time I fear that stat goals could be misinterpreted all the same, because they are actually saying “hey, you know, as an info, 95% of the best combinations lie here!” but not “you should always definitely aim for this!” because of actives, set bonuses etc.
However, if you decide to give them such a central spot in the presentation of the best in bags/slots, I think they deserve to be presented in a way that is rapid to compare to the actual pc stats, e.g. in the way ddcorkum proposed, otherwise the comparation will always be clunky because we’ll need to mouse-hover on the right column to see our stats, and see by eye where we are standing on the chart.
I see your point, that creating an hit-miss mark on the chart could make the false impression that best in bags combinations that miss the colored zone are “wrong” or “bugs”, but I think the community would be best served by understanding that “all truth lies in simulation” and that stat goals, even if more comprehensive than stat weights (i.e. “truer” outside of a neighborhood of your actual stats), are NOT the only side of the coin, while still giving a more complete presentation to data.
I completely agree with ddcorkum. Please add red markers of where current gear configuration sits. Second, not sure if another poster mentioned it. I also reference warcraft logs. Would it be possible to put that information into the DR (data repository) then mine it for the stat averages of say the top 10 or 20 percent of dps for that class/spec? Put those averaged results say in Purple (use a 1% percent margin?). Now we will have the range, the red (your current), and the (purple) the average from warcraft logs of the top dps configuration stats?
We are working on a big UI/UX overhaul and we will be revisiting how to display stat goals.
Including stats found on top log parses has a lot of problems associated with it. Gearing at that level of play is subject to a lot of trendiness and confirmation bias. We are trying to provide a completely data driven gear solution that is trend-agnostic and based solely on math and statistics, not feelings. We think this will ultimately show players that there are a lot more viable options for top level gearing than they have been led to believe.
Unfortunately we are fighting against many years of players being trained to use simple stat priorities and to only consider local maximum solutions to the gearing problem. The stat graphs we have now are a step in the right direction, but we still need to find a better way to explain what the optimizer is doing.
I have seen a lot of people want this indication of where their stats fall in the range. Presumably so they can see how their stats should change to be more optimal. The reason we have not done this is because it indicates that we didn’t achieve our goal of explaining the gear solution. If users understood what we were trying to show, they wouldn’t ask for this. The feedback is very useful for us going forward as we work on the next iteration of the site.
for me with yellow fives explanation the chart itself is good to understand.
Where i am always not so sure is as well the topic about “typical class stat” vs. current optimization.
example: my subtyl rogue gets the following after BiB: (ilvl918)
with this char, i only go lfr … but anyhow, i have similar results with my hc chars.
If i choose BiS: (only ilvl922) then it completely changes to mastery, even just 3 ilvl higher
clearly understood, different leggies impacts everything, but from major crit to major mastery is quite different, no?
$55;EU;Ysera;Echeyekee;Mörsertrupp;4;1;110;6:716,1:748;3;.s3;27;1113113;851,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,483,37,199,1,1,1,65;1,4,4,4,4,4,4,4,4,4,1,1,1,1,1,1,1,1,4,1,1,17,1;141255b1572b3528b3573,151013b3194b3337b3396,139253b1532b3528b3573;.q3;128476s16b743t163x141255m1572m3528m3573y9758n3194n3337n3396z-11760o1532o3528o3573c111;3s17t164;1751s12b-143b79b10b2679t166x12327c1e5427;6s2b-2768b79b10b1062t60x0c1e12;6540s6b-189b1966b45t168;293s7b-1762b1759t169;5097s13b-2048b1815b236t170;4503s8b-114b111t171x4y0z0c111;500s5b-2088b1855b227t172;1s15b-2097b2061b36t173e-4;1s10b-2092b1864b228t174;3s3b-2097b2061b36t175;593s11b-2077b1849b236t176e-8;1917s9b-1995b104b1655b237t112;1506s14b-2091b1854b228t177;398s1b-2964b1088b1720b200b1t178x-4c1$g\141255\141255\@g\151013\151013\@g\139253\139253\@g\151580\151580\200 CriticalStrike@g\151584\151584\200 Mastery@e\5427\128541\191013\200 CriticalStrike\124442=4,124440=35@e\5439\128553\190894\Mark of the Hidden Satyr\124442=15,124441=12@e\5435\128549\191021\200 Agility\124442=8,124440=20,124124=2
$55;EU;Ysera;Echeyekee;Mörsertrupp;4;1;110;6:716,1:748;1;.s1;25;1133121;331,15,930;1,1,1;152290b1472b3528b3614,151011b3174b3336b3396,;.q1;128869s17t161;1s16b741t55x152290m1472m3528m3614y-1279n3174n3336n3396c111;1360s12b-141b79b10b2679t166x569c1e5427;6s2b-2768b79b10b1066t167x0c1e12;6540s6b-193b1966b45t168;293s7b-1762b1759t169;5097s13b-2048b1815b236t170;4503s8b-114b111t171x4y0z0c111;500s5b-2088b1855b227t172;1s15b-2097b2061b36t173e-4;1s10b-2092b1864b228t174;3s3b-2097b2061b36t105;593s11b-2077b1849b236t176e-8;1917s9b-1995b104b1655b237t112;1506s14b-2091b1854b228t177;398s1b-2964b1088b1720b200b1t178x-4c1$g\152290\152290\@g\151011\151011\@g\151580\151580\200 CriticalStrike@g\151584\151584\200 Mastery@e\5427\128541\191013\200 CriticalStrike\124442=4,124440=35@e\5439\128553\190894\Mark of the Hidden Satyr\124442=15,124441=12@e\5435\128549\191021\200 Agility\124442=8,124440=20,124124=2
I would need the string from your addon, not the export from the website in order to recreate this. I’d have to see what legendary items are being chosen to comment on how the stat priority could change like that.
This is a good example of what we are talking about, though. It’s feasible that with your current gear, you want to get more crit, but making a change to something like a legendary item or set bonus can shift your desired stats significantly.