As to the point of the thread, the system is not perfectly elegant but it’s also not that bad. A little practice and you get used to it.
Activating heavy linker by a hold is way better than activating enders by a hold. If you’re going to let half of your beginners forget about a given technique, picking heavy linkers instead of enders is the obvious choice.
The only other way I could see it done is shadows are always 3P, and then enders could be 2P or 2K. But this is annoying in its own way, would increase the possibility of bad inputs (because pressing 2 or 3 buttons at the same time should be minimized), and SF players would be irritated that “EX moves” aren’t input by the now-standard 2P or 2K method. As it currently stands, if you forget about the hold mechanic, it’s easy to play the game without heavy linkers, because everything else feels in its correct place. The worst thing that will happen is you accidentally get ender when you don’t want, but at least it’s still a combo.
Also, pretty sure activating enders by hold is not easy to program. The reason you can disguise heavy linkers in a hold is because they can design the game so the startup 8 frames (or whatever) for a specific linker of all 3 strengths looks identical (and I do believe this is what they do). They then do not allow break windows during linker startup (again, this is what they do). The game does not feel disconnected here, because the game can just “animate the linker”, then make up its mind 8 frames later if you are holding or not and transition into the appropriate strength. This has pleasing gameplay aspects too, like letting ADs be breakable immediately but linkers not, which spices up the combo breaking game very nicely IMO.
If enders were hold, they would have to… play the heavy linker for 8 frames, then kara cancel into an ender that doesn’t have the same starting animation? And you would have to make sure that loooked nice for all 4 ender levels? The effort required to make it look good is much higher.
I’m still in favor of having heavy linkers use HP/HK. Perhaps the team could create a test build with the new controls and see how that changes any opinions.
Makes sense. I also agree that the current solution may be the better one given how the mechanics work.
I’m curious about how many frames the game uses to register a heavy linker. Now that the game has 5 frames of buffer to input a counter against a frame 0 breaker, that means if your 8 frame estimate is correct that leaves 3 frames to detect a linker’s strength.
About enders, maybe if they went with that they would then apply the solution you mentioned above to enders. They would have the same startup frames as the heavy linker and there wouldn’t even be a need to make the medium and light linkers to also share this startup animation, although your idea about linkers being harder to react based on startup animation is good enough that they would probably just make all versions of a linker share startup animation.
You are diminishing the value of baiting breakers to counter, heavy linkers are easy to bait breakers since they are so easy to break on reaction - after two or three counter breaks the opponent will probably not break them anymore. They also aren’t as obvious a bait as heavy ADs. Also they make medium and light manuals easier and harder for the opponent to predict and break.
Well, this is only true of manuals, because manuals are the only thing that you can’t break during startup.
It’s not true of linkers, which means you CAN counter break linkers immediately during startup if you want. So the buffer doesn’t come into play here and wouldn’t impact how long you have to hold the button down to get heavy linker.
I don’t mind having my claims questioned – I don’t think anyone is owed the benefit of doubt on making any pseudoscientific claim they feel like, much less myself – but it’d be nice if people could phrase their concerns in a way that gives me a little more credit than just outright assuming that I forgot that counter breakers exist. I was around in S2 rebutting people who were so defeatist about the combo system as to think that opener → once-chance → launcher → sweep was the only engagement they could risk, after all.
But okay, let’s do some simple modeling. Say, assume in a Jago mirror that player 1 landed an opener straight into a heavy linker (putting aside whether starting with a heavy linker is a good idea) for ~10%. We’ll assume player 2 can comfortably react to the heavy linker and didn’t make an immediate guess break attempt, so that the only important decisions on the table are whether player 2 goes for the heavy breaker, and whether player 1 goes for the counter-break. We’ll also assume that both players are well aware that this mind game is taking place, make a best effort to remain unpredictable, and have analyzed the stakes in advance so that they can doctor their tendencies to maximize their expected payoff from this interaction.
So, what are those stakes? Two lines aside yields four possible outcomes:
Both players let it rock: the combo continues. Player 1 might get in a heavy manual, player 2 might attempt a guess break (and most likely lock out on it), but all in all this isn’t much better than the follow-up to a medium linker (mainly because guess-breaking after a medium linker isn’t very good either), and getting a heavy linker past player 2 doesn’t really raise the expected damage of the combo substantially. This gets into computing the expected damage of a typical combo opening (yes, factoring in breakers and counter breakers throughout), but in lieu of a full solution on that front (hopefully I’ll have time to finish this soon!), the folklore figure of 25% will have to do.
Player 1 lets it rock, player 2 breaks heavy: the match reverts to a situation which doesn’t really benefit one player very much over the other, but player 1 got in 10% before the break. This is going to vary from matchup to matchup (e.g. zoners generally gain a lot of space off a breaker, which is usually worth more to them than their opponent), but we’re looking at a Jago mirror to keep things simple, so 10% will be the net outcome here.
Player 1 attempts a counter breaker, player 2 lets it rock: with a good reaction and a bar of meter, player 2 could make the punish hit for ~20% before anything breakable happens. But I’m going to assume this is taking place at the start of the match before either player has meter, and I’m going to place the expected damage off the punish combo at 25% (there is room to argue that this number should be higher), which less the 10% player 1 took before dropping their own combo is -15%.
Player 1 attempts a counter breaker, player 2 attempts to break heavy: I’m going to lift the number for the optimal meterless counter breaker punish from Infil’s Jago page without doing anything to adapt it for our assumptions, because it’s pretty close to what player 1 should be getting optimally. That’s 50%.
Now that we’re aware of the stakes, what might be a good strategy for player 1? Well, the minimax principle suggests it might be a good idea to pick a strategy for which player 2’s best response is as bad as possible. What does that mean? Well, assume player 1 chooses to counter-break with probability p in [0,1]. Then the expected outcomes for player two’s options are
Player 2 lets it rock: -15p + 25(1-p)
Player 2 breaks heavy: 50p + 10(1-p)
First, a quick check for degeneracy: if player 1 never breaks (p=0), then player 2’s options yield 25% and 10% respectively; and if player 1 always breaks (p=1), then player 2’s options yield -15% and 50% respectively. Basically, player 2 doesn’t have a dominant response: there are values of p for which letting it rock is substantially better, and likewise values of p for which breaking is substantially better.
What player 1 wants, however, is a guaranteed amount of lifeswing no matter which strategy player 2 picks, which entails finding that value of p somewhere in the middle which makes player 2’s strategies as good as each other. So, a bit of high school math:
break heavy < let it rock
<=> 50p + 10(1-p) < -15p + 25(1-p)
<=> 65p < 15(1-p)
<=> 80p < 15
<=> p < 15/80 = 3/16.
Somewhere between 1/5 (=3/15) and 1/4 (=4/16). So player 1 should be counter-breaking less than a quarter of the time, lest the “let it rock to clean up on a whiffed counter breaker” strategy gets too strong.
But what does this tell us about the expected lifeswing on choosing to counter break with probability p=3/16? We only have to plug p back into one player 2’s strategies, since either one will give the same outcome:
Further, if player 2 takes a similar approach and finds a probability q between 0 and 1 with which to attempt the break, then by the minimax theorem, they can find a q (I’m not going to do this calculation to find q, but you have all the information you need to do it yourself if you want, and I’d expect q to be ~1/2) for which neither of player 1’s options does better than 17%.
Mind, if you go to the trouble to figure out the payoffs for light and medium linkers – which isn’t so straightforward to do by hand, and is probably best done with linear optimization software – you’ll get expected payoffs north of 20%, probably around 25% now with the S3 changes to combo breakers. You’ll also find that heavy auto doubles fare better because, whilst presenting a similar-looking scenario, heavy doubles actually reward player 1 well for getting player 2 to let it rock, whilst heavy linkers don’t. (If you want to see a situation like this where player 1 fares really well, look at TJ’s heavy auto double game and factor in barrage. Best combo trait in the game.)
None of this is to say that heavy linkers should never be used – when I’ve modeled the situation it hasn’t turned out that you should never use them, in practice your opponent’s ability to defend against them is going to vary, and hey, the model above isn’t perfect – but I take it as a strong hint that medium and light linkers should be your bread and butter, and heavy linkers should be used sparingly or provisionally.
Making your opponent take a 50/50 guess break attempt is fine: if you have 10-15% damage already, and you can find 40% on the lockout, then you’re averaging over 25% if they attempt the guess break, and if you factor in things like meter, timing lockouts, etc, that number gets higher. If you go in and solve the situation, you’ll find that an optimal defender rarely attempts a 50/50 guess-break at all.
If you really want to use HP and HK for heavy linkers, you can actually just use Combo Assist. It uses this convention, and since you can customize it so much, you might be able to set it to a “linker only” version of CAM that gives the effect you’re looking for.
Re: the new buffer window and linkers, Infil is correct. You’ve always been able to counter break linkers on startup, so the new buffer doesn’t come into play at all.
@Fnrslvr - always interested to see the analytics/optimization toolsets you apply to KI. You’re like the moneyball of Killer Instinct.
I’m an engineer, so my head tells me that long-term, you are generally probably right on most of your premises regarding combo theory. I do think a (maybe big?) stumbling block though may be the assumption of “optimal defender”. In practice, very few players indeed approach the combo game optimally, either on offense or defense. And we see pretty prevalent 50/50 guess breaks kind of all the time, even (and especially) at high level. Light linkers, medium/light manual breaks - there’s actually a lot of combo-game guessing that does wind up playing into matches, precisely because real players do not (currently) play very optimally within combo. But who knows - that may just be my gut responding to the premise that I’m playing the combo game all wrong
Sorry if my response seemed rude. It can be hard to get away from the “just get good” mentality.
Mind games and hard reads are a big part of why I enjoy fighting games so much, and Killer Instinct specifically. I wanted to restate their importance because I know one read can often make or break a match.
Random thought - I’ve recently begun wondering just how much resets (command throw and anti-jump out resets specifically) affect the combo game. Basically, it feels like the more I force an opponent to think about these things, the more likely I am to be able to get away with otherwise-reactable options like heavy linkers and med/heavy AD’s. By overloading (or trying to anyway) the opponent’s capacity for multi-tasking, I feel like I’m able to “get away” with more damaging options in the combo game without having to resort to counter breakers.
Against the very best (optimal) players of course, this theory begins to see decreasing returns. But even here I’ve found that it can often yield results (or perhaps that’s simply the optimal player refusing to bite on obvious counterbreak bait?) Basically, I am just wondering if there is an analytical way to describe this attempt to overburden the defender in this manner, or perhaps to just add another damage option to the aggressor’s suite of options within the combo game. The latter seems like it’d probably be the better way to do it now that I think?
As a new player, I can definitely say that heavy linkers are massively unintuitive. In my game, I barely use them because it still requires such a conscious decision… which I guess is actually probably a good thing.
In any case, I have no idea how they could intuitively solve them though.
While I do agree with your numbers (nice work by the way) I originally had a problem with this comment:
I’ve typically found in any modeling I’ve done that they’re almost not worth the trouble…
I don’t believe this to be true and I also don’t thing your numbers support this. Even you eventually said you weren’t implying HL should never be used so maybe it was just bad phrasing.
While in an optimal scenario you can reliably just use ML and LL and do well I don’t believe KI is the kind of game where you can sustain optimal play for very long. Also I believe @STORM179 has a point when he wonders the effects of resets in all this math.
For example, with Kim Wu we can dragon cancel a HL into a manual having both the benefit of full range of manuals and baiting lockouts with a HL. How does that affect the mind games? And how about instinct cancels?
Anyway, good analysis on HL usage. I believe KI still has a long road before we start seeing people playing at a truly high level where this kind of data can make or break a match.
[quote=“Fnrslvr, post:19, topic:10402”]heavy linkers aren’t even all that good…
I’ve typically found in any modeling I’ve done that they’re almost not worth the trouble
[/quote]
Really? You’re only playing 2/3s of the game then. Your opponent is more than likely using all of his tools which puts you at a disadvantage anyway you slice it.
So, this shouldn’t be an issue for the offense per se, because the minimax principle seeks to find a strategy that guarantees a baseline payoff against all possible plays the opponent could make. So the strategy you end up solving for doesn’t just perform well when your opponent is playing “correctly” (which is one of the pitfalls of solutions concepts in games that aren’t directly adversarial and/or games involving more than two players), but performs indiscriminately well no matter what strategy your opponent opts for. It’s a pretty good candidate for the definition of “conservative play”. (Unlike one-chance break play that we saw a bunch of in season 2, which is more like just plain paranoid play.)
Mind, one of the disadvantages of taking this approach, is, say your opponent is doing something dumb – like always breaking your heavy linkers. The catch-all optimal strategy we solved for is still going to insist we counter break a quarter of the time and accept our 17%, but with the “knowledge” that our opponent doesn’t know better than to just always break, we could just counter-break every time for an "expected’ payoff of 50%. Here we see that the optimal strategy is truly blind to the substance of our opponent’s play: unless they’re playing badly and opting for lines that they should never be using at all, optimal play won’t improve your outcome if your opponent’s habits are skewed one way or the other. In order to capitalize on your opponent’s sub-optimally skewed tendencies, you yourself have to leave the relative safety of the optimal solution by adopting skewed tendencies which counter your opponent’s skewed tendencies – which opens up the possibility of your opponent reworking their tendencies to counter your counter tendencies, etc.
With this in mind, models of optimal play are probably best thought of as a kind of baseline: they don’t tell you exactly what you’ll see in practice (because even in a Bayesian sense, your opponent breaking every time thus far is mounting evidence that they’re likely to break again), but you can expect mind games to move around in the space surrounding optimal play, and importantly, modeling for optimal play is probably the best way to assign a “value” to a strategy in the long run.
I may’ve slipped a contentious assumption past when I said “both players … make a best effort to remain unpredictable” – to my understanding, what research has been done here seems to suggest that people are kinda bad at flipping coins in their heads. Patterns emerge, often without the mental coin-flipper noticing, and the person tasked with predicting the string of flips is often capable of picking up on those patterns and performing better for it.
But it seems to me that there’s little reason this wouldn’t work both ways: what’s stopping the mental coin-flipper from getting a read on the predictor? Truth is, I don’t know nearly enough about this kind of research to really say. But if we were prepared to ascribe too much to this effect, or admit ignorance in the face of it, I don’t know that you could have balance discussions about concrete numbers like damage and frame data of a move and whatnot either, because a good series of reads can overcome that stuff too.
I don’t expect this stuff to invalidate my modeling – if heavy linkers come in at substantially lower expectation than the other linkers with my assumptions, then I think that’s strong evidence that something’s up – but I guess I’m open to well-considered counter-arguments.
S’cool, I get where you’re coming from.
Caring about the mind games is what spurred me into developing this approach in the first place. I don’t think you really need to model much to know that the season 1 combo system was pretty degenerate: arbitrary manual opportunities off of a lot of linkers, always level 5 shadow ender cashouts, etc. Basically no reason to do anything reactable, or really any incentive to weigh the risk-reward of anything, just clearly dominant strategies all round that ignored most of the game’s combo systems. If you’re a fan of mind games and reads, you may be able to carve out some reads around the edges of season 1’s combo system, but it certainly wasn’t fertile grounds for reads to grow in.
There were a bunch of people a while ago (including some competitive players who probably don’t deserve to be named) who argued that the season 1 state of affairs was better, and that made it important to me to examine whether
season 2 exercised the range of KI’s combo systems in a way that season 1 didn’t; and
season 2 combo damage was still high enough to reward the offense for getting an opening.
I went far enough with some ballpark figures and simple modeling a la the stuff we looked at before, to convince myself that both of these were indeed the case, and I think as we watched competitive play adapt over the past year we saw some of that bear out in practice. But the robust way to go about this stuff is to write a program that considers a few hundred thousand scenarios like the one we solved above, parameterized by things like potential damage and KV, that all feed into each other, and then line them up and “backsolve” them. I got started working on this, but I just haven’t found the time to finish it yet.
I’m probably not going to model resets even if I get some of my more ambitious stuff off the ground, but I definitely think they push expected lifeswing upward, probably pretty uniformly. (i.e. I can’t see why they’d benefit any particular stage of a combo in particular more than any other, aside from not helping the reactable break/counter-break mind games, since you obviously can’t reset in the middle of the reaction game we just dissected, if that makes sense?)
I generally took them as further evidence that all the complaints in season 2 about the combo system not rewarding the offense enough were overblown, since even without factoring resets in (and a number of other hard-to-model things that almost always favoured the offense), expected lifeswing off an opening still seemed to be over 20%. Since a major goal back then (and probably still now to some extent) was to lower-bound expected damage to show it wasn’t anywhere near as bad as some hyper-risk-averse players would’ve had you believe, I took this as a good thing and left it at that.
That’s because you probably are.
Overloading your opponent’s “mental stack” is yet another thing I think is super-iffy to model, but that probably favours offense almost entirely. If you’re super-interested in this, you could do some reading on mental chronometry, which is the study of human reaction times to different types of events. Apparently there are models which, when you’re looking at n possible options that you need to look out for, posit the usual up-front delay, plus an additional duration that scales logarithmically in n, as though something involving Shannon entropy is going on, the brain is feeding the stimuli to a well-optimized decision tree or something. But this stuff has a heavily experimental aspect to it, and I can’t see myself modeling it credibly, so it’s just another one of those things that compels me to stick to the simple model that confidently lower-bounds actual damage, and hopefully gives you a good idea of how different strategies look relative to each other.
I think my phrasing back then is along the same lines as what I was saying at the end of that modeling post, and also what I’ll say now: heavy linkers should be used sparingly or provisionally, and you’re not really hurting yourself if you don’t integrate them into your gameplan at all.
As to my numbers supporting my claims, I absolutely think they do: light and medium linkers do so much better than heavy linkers in expected lifeswing terms that a ~50/50 combination of the two generally outperforms any combination that gives substantial weight to the heavy linker option. The added ambiguity on the guess break just doesn’t make up for the cost incurred by worse outcomes when the opponent chooses not to break.
Kim Wu with dragons is an edge case that applies to one character, so you’re saying that…there are exceptions to the model so the conclusions drawn from the model don’t carry any measure of truth? I think you’re either taking my claims in a very absolutist direction, despite the presence of qualifying terms like “almost” and “typically” that offer plenty of wriggle-room for exceptions to go off any do their own thing, or you just think that a simple model that doesn’t account for exceptional circumstances cannot get a proper handle on overarching trends.
Instinct cancels are a better example since they’re more universal – but so are things like meter (meter on either side changes the numbers, because “losing” the interaction gets you blown up harder), amount of health left on each side (both for guts scaling and because the “value” of an interaction becomes less about expected damage and more about things like Pr(player 2 dies | strategy E) as health becomes more limited), positioning, matchup considerations (there’s probably a good argument for being more a risk-averse aggressor in-combo if the defender is a zoner), demonstrated player limitations (e.g. if the opponent seems incapable of reacting to heavy linkers), etc. Instinct cancels do add another strategic line to the model, but you get ~2 cancels a match and at some point it becomes silly to fret over every situation that access to an instinct cancel can blow up. If it interests you, you can model it – but again, like all the complications I just dreamed up off the top of my head, these are complications which detract from the goal of using a simple model to get a grasp of overarching trends.
This is like suggesting, as someone once did before in another thread to me, that there’s always a 50/50 chance that your opponent will counter-break when you try to break. You need to consider the risk-reward, weigh in the stakes, realize that your opponent’s interests have considerable bearing on the probability that they will do something. Whether heavy linkers get their one third of the pie as you’d suggest depends on how their expected reward stacks up against the other two options, and I’ve already spent a lot of characters arguing my position that they don’t.
EDIT: if your point is that heavy linkers are great because someone who thinks of them as a free breaker every time just gets wrecked repeatedly, then that’s something that has been discussed above.
I think I got confused with what you meant by “almost not worth the trouble”, which to me isn’t the same as saying to use them sparingly.
Yes, your model shows that in typical scenarios you can just not use them and be perfectly fine, but you also acknowledge that sometimes you have to deviate from optimal play because your opponent is deviating from optimal play.
Anyway, I think we are just discussing semantics here. Fact is I do believe your model works.
No, that’s not what I said. Again, I do believe the model works. This was just me wondering how the extreme cases affect the model, not me trying to invalidate the model.
But as you said, such cases may be just too complicated to model reliably.
