Frequently Asked Forum Questions | ||||
Search Older Posts on This Forum: Posts on Current Forum | Archived Posts | ||||
re: DMR Study Update (new spreadsheet) | |
Posted By: RC Master | Date: 1/27/12 6:15 a.m. |
In Response To: Re: DMR Study Update #3 (aka for Hoovaloov) (Hoovaloov) : Ok, I can respect that. Cool. I think we've delt with the next few paragraphs well enough on TTL so I'll skip over those. : Yes, I was saying my conclusion is stronger because my frame of reference was
Right ok, I understand now. : Also, I talked to Joel quite a while about his equations. ...
I've thought about it a bit more, and normal distributions should be a good fit for the pacing curve, and good fits for each individual spamming spike, but they'll never be a good fit for whole of the spamming results because of the gaps in between where it's possible for a spammer to fire. So, if you wanted to use normal distributions to model the spamming probability, it'd have to be the sum of several normal distributions centered around each peak. Which gets more complicated... : I suppose that wasn't very clear, since I don't think Google Docs uploads
Ah right. Well if G Docs has macros I've no idea where they are anyway! Thanks for explaining it, I didn't really want to comment on that part until I understood what you were doing. : What do you think? Since the order of the wins don't matter (you just count the wins in each set), you could have gotten away with only randomising the spam wins order. : It's like we have two decks of 25 cards, one for pace and one for spam. Then
Problem is, this isn't a good model of what actually happens. A closer analogy is:
See, the way you've set it up, if I draw a card that says '49 frames', that stops me from drawing that card again in our 25 card run. When in actuality, when I'm firing my DMR, I could randomly draw it multiple times to beat many of your cards. Yeah, maybe your method doesn't end up too far off the mark anyway, but there is a much more elegant, straightforward way of doing it and we don't need 1,000 sets of 25 card runs, we just need to look at 625 possiblities (25 * 25). Take a look at this Google Spreadsheet. Thats not even the most elegant way of doing it - since there are duplicates in every sample, you could just have the unique combinations and then also have the frequency of each card in each of the pacing and spam sets. Doing that would also lend itself to adding more test data more easily. You can see an attempt of this on the page 'Unique Possibilites'. But unfortunately I can't seem to figure out a way to combine the frequency weights to get even basic stats out of it. I'm not done with this spreadsheet yet - some niceities aren't all there, and there are a wealth of derived stats you can do which I haven't even begun with. Regardless, this spreadsheet should still get the point across that you don't need to worry about random numbers and samples of potential DMR battle series or whatever since you can actually enumerate all the possibilites with the given samples from both pacing and spamming style of shooting. Lag ToleranceIf you're looking at that spreedsheet you're probably wondering what all the blue boxes are. This is an extension of what I was talking about at the end of this post about pings in games. I believe this is quite an important issue to discuss and include in the analysis since Reach is a game which is primarily played across the internet and that mode of play is most player's frame of reference - does the fact that your pacing killed the other guy matter as much if he simultaneously killed you with his spamming? I don't think it does since your pacing clearly gave you no advantage in that case. At the very least you should model strict draws - where both spamming and pacing kill in the same number of frames - as one would assume that even when played locally or LAN this theoretical situation would result in a mutual kill and give no real 'win' to either player. I noticed you didn't do this in your simulation. Given that we know that there is some lag tolerance above that, I think it would be a mistake not to include higher frame differences as being 'draws'. I suggested 3 frames (100ms) either side (pacer could be either host or client) as a start and that is what I began to include in the first spreadsheet. JonnyOThan thinks thats even on the low side for the amount it may actually allow. I have personally seen a 5 frame difference between when I fired and when the opponent fired and we both ended up dying, just a few days ago. We could have been both clients, granted, and that is of course going to make the lag between the two higher in any discussion. I'm sure if we paid attention we could find examples higher than 5 frames anyway. So, 3 frames is not unreasonable for a 'draw' consideration (no benefit to either side) and this is what I modelled on the 'possibility matrix' sheet. You can see that things start to look a bit different when you take this into account - as I suspected when I first saw the data. If you look at the absolutist stance where a difference of 1 frame is a 'win' then sure, you might still be able to conclude that in theory the 85% Bloom DMR rewards 'pacing' more than 100% bloom did. However, the reality of the situation is that there is lag, and when you start to compensate for it, you end up with a lot of draws on the tables. I'm not sure that such an increase in no-one winning (or who wins being entirely up to lag) is desireable. Wasn't randomness and inconsistency one of the complaints? One of the things that 85% was trying to solve? It seems you just shift that randomness of the shots into more randomness of who will win - perhaps no-one will, perhaps it'll be decided on lag. The second sheet in that doc - 'Win Margin' - goes into more depth about this. It shows the number of frames of difference (negative is better for spam, positive is better for pacing) and then colour codes it for three levels of lag tolerance and absolute draws.
Even the 9 frame draw represents about 150ms ping between host and client (300ms each way) which can still happen (though of course I don't know if the lag compensation allows that, and yes JonnyOThan did say it was more complicated than that). Or if you consider players on two different client boxes, you might consider that as 75ms ping between each client and the host. Let me explain that example: Client 1 fires a killing blow on Client 2.
So on Client 1's Xbox, it looks like they died from a player they should have killed 9 frames ago - but in the context of lag, it's still fair, because Client 2 thought they were still alive when they fired, and it's hitscan to make shots reliable anyway! If you don't want to agree to the lag compensation thing right away then please, lets discuss it. In the mean time, feel free to take a look at the page 'Absolute (no lag)' - there the draw tolerance is set to zero. The numbers there are close to what you got anyway, but this is a much better way of getting there and should be easier to understand too. : If you're asking what the threshold of detectibility is, or how much of an
: Let's say we play a game to 25 kills, a 1v1. We each will only fire our DMR
Ok, let me talk about why I don't think the results at short-range (in any of the analyses) are necessarily a good thing here since you have an issue with that. Firstly, look at it with lag compensation factored into it. There is a greater direct overlap between between the possibilities, and the question of 'who wins' is muddier. But both of them have less outright wins in 85% than 100%. Secondly, even without lag compensation, the 'what strategy should you use' gets less clear. In 100% bloom, the smart man's money - the person who understood bloom and what your chances were - would bet on spamming over pacing from your data (figures quoted at the end of Update #3) and techniques. In 85%, smart man is still betting on spamming, but the odds are reduced. This means there is less difference between the strategies - which is obvious because Spam is still ROF locked, and pacing can fire faster than before with full reset. Instead of 'rewarding' the pacer, what you've actually done is punish smart man - the person with the greater knowledge and insight is now going to lose individual fights more often. Smart man would of course switch to using pacing at medium and long range. But again, when you factor in lag, smart man(now pacing) is punished in 85% with less outright wins. Yes, smart man/pacer also doesn't outright lose as much as either, but lets look at the score difference in a game to 100 kills: Medium Range
Score 100-64 -- pacer wins by 36 85% bloom
Score 100-69 -- pacer wins by 31 Spammer gains 5 kills Weird huh? I'll do the same thing for short and Long range: Close Range
Score 38-100 -- spammer wins by 62 85% bloom
Score 100-50 -- spammer wins by 50 Pacer gains 12 kills Long Range
Score 100-18 -- pacer wi50/88ns by 82 85% bloom
Score 100-22 -- pacer wins by 78 Spammer gains 4 kills For a standard game to 50 kills you can half all that - Pacer gains 6 in Short Range, Spammer gains 2 at mid-range and 2 at long range. So at short range, the increase in winning percentage is noticable, but smart man's money is still betting on spamming. At medium and long range, the difference isn't much, but is actually in favour of spamming. : ---------------------------------------------------------------------------- : A spam/max ROF fire shot is 12 frames in both 100% and 85% bloom. For the
From the bloom you get after firing 4 shots at max ROF? It's really just 3 extra frames? : Also, add the 2 frames it takes the last shot to register. So that
They certainly sound reasonable but I haven't looked at the reset rates in any great depth myself so I'm just taking your word for it on that front. They also sound insanely quick - particularly the 100% one compared to the spam times. Like, seriously. Go look at what that implies in short-range... thats... insane. : 85% bloom takes away the one shot "cushion" the spammer has. He can
But the spammer doesn't have a very good chance to hit on the 5th and 6th shots anyway. Look at where the quartile distribution is. : Therefore, 85% bloom rewards the accuracy of
Conversely, if the pacer misses and the spammer remains on-target, suddenly the spammer has a good chance to win on the 6th shot (which was my point) - again, quartiles. : You were pointing out very specific 1v1 encounters to make a point, I was
Hmm. I guess I just had a problem how it seemed to be that you were just quoting the 6 pacing win percentage figures as 'the whole story' when this data actually says a lot more than that. : The second part of the hypothesis was that the best way to fire
In TU or default? JonnyOThan reckons bursts of 2 are optimal for both weapons. At least in terms of bloom, in 100%. I tried it a little with the DMR in some TU BTB and it still seemed to work pretty nicely until I started freaking out when they were firing back :P : Sorry for the crazy long post. I'll upload the maps and gametypes for you
Email, XBL PM, message here. Whatever. I might not have time to get around to doing testing for a while though. I have masses of Uni work and even more personal projects to do. Constructing the new spreadsheet took a while as well 0.o
|
|
Replies: |
The HBO Forum Archive is maintained with WebBBS 4.33. |