Hey, any #guildwars2 players on the fediverse into data science and interested in helping me with a little (as in specific, not little as in "not a lot of data to parse") project?
@pixel following with interest because burst is such a weird matter.
There was some chat about this in the Help Desk thief channel recently.
Do you have ideas for what you'd consider a useful burst window? I have thoughts on this, but it is probably just going to make things even more complex lol
@Lucinellia the more I think about it, the more I think measuring burst as a function of "time to damage threshold" rather than consolidating it as "damage per second" over a window helps with general legibility. If I tell you "bladesworn takes 12 seconds to do 500k" that means more than like, "bladesworn has 76k DPS at 30 seconds" or whatever. You don't mix measurements too, that way burst is time and bench is volume, which helps public legibility.
@pixel Yes, that makes sense. It is a handy metric for sure, and one that isn't considered sufficiently in balance.
Burst is mainly important to the wider player base in HT CM in my opinion so I have always considered burstiness to be the ability of a profession to contribute considerably to the giants and goliath / obliterator moments of the fight.
That is where it gets messy - giants are 3 targets, obliterator and goliath 2. So the burst that matters is also linked to cleave...
@Lucinellia yeah I think cleave is just like, something worth adding as a positive on SC, i.e. if a build has exceptional cleave mention it, or if it's restricted to single target then mention that as a negative, but time to x damage can be taken in concert with those to form a bigger picture
@S_H_K so, different fights have different windows of opportunity for DPS, and builds that can consolidate their damage into super tight windows are "burstier" and tend to fare better in content where you really need to do damage in short windows. I was going to do some work to figure out how to compare data for DPS done on a benchmark (i.e., in a vacuum) and find ways to plug that in to some approximation of a "burst window" -- whatever seems like the most accurate way to represent that. If that makes sense
@pixel
Alright I have a rough Idea of the game (MMO it seems) I'll have to check more about how is the game but there's some questions.
1)Is there some element system that gives bonus damage or penalty to damage?
2) Is there animation cancel in the game
3) You want to explore the possibilities of combo routes or just compare general stacked damage?
attached are four images of best-case DPS (average dps over time charted) for four different builds, basically i want to see how best to display a build's outgoing damage over a period of time that best represents a realistic context for what burst looks like across all encounters, if that makes sense
@pixel
I found the charts and Damn my eyes are still spinning with all that data dump although I'm still unsure how to interpret that thing they seems to call dps to the damage output maybe I misunderstood that. Anyway going back to the questions.
We would have a time axis and a damage axis for sure don't ya?
If we are only working with one build "the third axis" is the different encounters.
I see there are factors like quickness and some other things they use interactive elements to show how they influence we would use a best case scenario or represent a rough delta of what's the DPS for the build? Or you want it to be interactive?
@S_H_K all boons and such can be assumed to have 100% uptime, most endgame play is maximized around that. basically I'm trying to solve two questions
how long is the average "burst window" across fights, which can be discerned theoretically by parsing logs from fights to see when they are and are not able to do damage
how much damage can each build do during that time
since DPS is really calculating "Damage (dealt so far) per second (of total time in combat)," are there better ways to display that data than "damage per second" -- is time to x amount of damage better? measuring burst as a function of time specifically?
@pixel
Oh dear I got a little lost Clarify me something
A)Builds one or many?
B)Encounters one or many?
C)We will be working with logs? Hope they aren't a mess
Looping back to the threaded conversation:
We'll have to define the burst window.
Ah so it's not DPS per se in this case
Yeah well if we define a burst window in time we can look at how much damage a build puts in that time though different encounters.
If we define the burst window as an amount of damage we can look at how much time it takes to reach it, and how reliable is to reach it.
@S_H_K im trying to design a framework to check any build, given a benchmark, against an approximation/model of most encounters' and how much burst they often expect. we would be workign with logs, but programatically sorting through them isn't my strong suit. I do know some friends that have worked with the log files and they're not horrible afaik, but I'm not sure how robust the documentation is
@pixel@S_H_K log parsing is more on the state machine side of things than data science, yeah? It really sounds like you're trying to track the state of the encounter and make certain calculations when the state machine says to
@neatchee@S_H_K no (i dont think), i want to find a way to create a model that's sort of an approximation of what a fight might look like SHOULD it expect burst, and then find a way to compare a benchmark's log against the expectations that a "burst window" would have and compare builds against each other under that context. if that makes sense
I'm not necessarily sure what the best approach would be, I figured this would be data science driven but I'm a layperson and may be wrong in that regard so feel free to correct me on what should/shouldn't be done
@pixel
Data science can be applied to a log colection no problem, I've done it loads of times. If the logs follow a clear standard it will be easier. I guess I'll have to see them to tell what we can get from them. @neatchee
A state machine is a coding pattern very common in gaming is usual to use entity behavior modeling which can be considered a subset of it. Logs are generally made when the entities interact or change states we could try track that but it will give more info about the code than the performance of a build. Or maybe you intended to say machine learning?
We're trying to read the logs to get data from gameplay
We not trying to calculate damage or use case scenarios they are already well established.
We want to explore how to tally the performance of a given build in general. Which is info the site given doesn't give really, it tells you best use case for a build to a specific encounter.
Given the last point with this visualizations we would be seeing a rough performance of the build in general and the variance (or delta if you will) of damage through different encounters and iterations of execution.
There's another site, https://gw2wingman.nevermindcreations.de/stats that posts logs on every fight. I'm not sure how easy it would be to scrape logs and determine "burst" windows mathematically, but this site has a ton of logs for a ton of encounters that you can use to model against, instead of the benchmark logs on snowcrows.
Add comment