## Stat Distribution Analysis: Part 2

May 31, 2018 Leave a comment

One week later and I think I’ve cracked this. Firstly, go back and read last weeks post if you haven’t already, and this old one on stat weights if you don’t understand why replacing stat weights is a good thing.

All caught up? Good.

My approach up until recently was to look at the DPS results for various secondary stat combinations and try to work out some sort of consistent pattern from one group (grouped together by the sum of secondary ratings) to another. This involved looking at average DPS values over multiple results where one stat had X rating, plotting the ratings vs DPS to look at trend lines, and other fun stuff. The “eureka” moment was when I looked at using each total rating set as stepping stones towards the next.

# Stat Weights & Other Stuff

As a quick recap, stat weights are generally used as an indication of what stats you should aim at based on sim results. These will often be loaded into Pawn so you can get a comparison of different items in-game. The problem with stat weights is that they are pretty specific to the current gear set & character setup, and what works for one player may be incorrect for another.

These days you can do “best in bag” type simulations comparing permutations of various items you already have to find your best combination. This can be done on Raidbots (via SimulationCraft) or on AskMrRobot (using their own simulation system), but this is only useful after you’ve acquired an item and doesn’t really help with the eternal question of “How much of each stat should I have?”

Some theorycrafters have tried answering this question via 3D charts with Crit/Haste/Mastery as the X/Y/Z axis and colouring the points based on DPS. This works, but requires running simulations for every combination of each stat at whatever interval you choose, and the graphs are still limited to that one total sum of secondaries.

# The Solution?

The goal I’ve been working towards is setting out an easy to follow table so players can say “I have a total of X rating, how should they be split and what will I want to aim for from there?”, look up their values, and adjust their gear accordingly. This can also support the secondary targets/thresholds/caps that players talk about when they see stat weights changes result in different stat priority orders.

For my test data I used 1000 rating intervals (as I’m using data for v7.3.5 as a test bed rather than on anything in Battle for Azeroth), and returned every combination of secondaries from a total of 6000 to 19000 rating.

- Find the best combination for the initial set (ie: 6000 sum total for Elemental is 3000 Crit, 1000 Haste/Mastery/Versatility)
- Add your interval to each secondary and re-run sims to find the one with the highest DPS.
- Add your interval to the combination returned in 2
- Rerun sims & repeat.

This gives me a “secondary ratio path” like that works as follows, covering a sum total from 6000 to 19000:

- Start with 3000 Crit, 1000 Haste/Mastery/Versatility
- Increase Mastery to 2000
- Increase Versatility to 2000
- Increase Crit to 5000 (two steps)
- Increase Mastery to 3000
- Alternate adding 1000 to Crit & Haste until you reach 8000/4000 respectively
- Increase Mastery to 4000
- Increase Crit to 9000

When comparing the DPS results of each of the above steps back to the absolute best result for each rating sum the biggest difference was 1.43%, with the lowest being 0% (the 3rd 0% result was actually at sum 17000). The difference over the last 8 sets was around 1%, which for something that will be used as a guide for the average player, or for alts, is pretty accurate. You can see the data/calculations here.

The next step is to get more dataz to see how well this works with smaller increments as well as comparing results between talent builds in a spec. I’m hoping that there will be close similarities, or at least differences that can be boiled down to behaviour of individual talents & interactions. At the worst there would need to be a single data set per talent combination.