AI 2.0 Discussion

https://forums.playcontestofchampions.com/en/discussion/392960/dev-diary-ai-deep-dive#latest .

It's nice that we have confirmation that Dr Doom is weighted to throw more heavies

Comments

Average_Desi

what change ai behavior :

Ai profile - ez for Arena , hard probably for rest
Champion Specific Tuning - chavez , mysterio, doom(heavy heavy)
Aggression, - chavez , mysterio (keep cornering you)
Mistakes - finishing combo on block or special on block
Effects(Taunt, Intimidate, Infuriate) - Classic Titania, Starky , gladiator....

how does light intercept , parry & Side step work ?

Venom and CMM are also ridiculous aggressive.

DNA3000

what change ai behavior :

Ai profile - ez for Arena , hard probably for rest
Champion Specific Tuning - chavez , mysterio, doom(heavy heavy)
Aggression, - chavez , mysterio (keep cornering you)
Mistakes - finishing combo on block or special on block
Effects(Taunt, Intimidate, Infuriate) - Classic Titania, Starky , gladiator....

how does light intercept , parry & Side step work ?

As the devs have often stated and the Deep Dive reiterates, the AI doesn't understand intercepts, or any tactic really.

But consider what an intercept is. How do we do intercept? We wait for the AI to come to us, and the moment our attack could actually reach the defender (depending on whether we are doing a dash forward or a light intercept) we throw our attack into that window of opportunity, hoping it lands before the AI can throw its own attack.

The AI doesn't understand "intercept." But it does understand "attack if the player decides to approach us and is within range." Since the AI is, as the Deep Dive mentions, extremely fast and precise, if the AI decides to hit you, it will throw an attack at you at the first instant such an attack would actually land. The AI doesn't have to hope, it calculates.

What does "throw an attack the moment the player enters into range of the attack" sound like? The AI just wants to hit you, but it won't throw a whiff. It will wait until the attack would actually land, and throwing an attack as the opponent approaches you to throw an attack themselves is what we call an intercept.

So the AI isn't trying to intercept. It is just trying to hit us quickly. But perfect timing on an attack while we are entering into range is indistinguishable from an intercept.

winterthur

Probably irrelevant but how does this kind of stuff sits in with the explanation given?

https://youtu.be/ubbniSiyzzs?feature=shared

Asher1_1

what change ai behavior :

Ai profile - ez for Arena , hard probably for rest
Champion Specific Tuning - chavez , mysterio, doom(heavy heavy)
Aggression, - chavez , mysterio (keep cornering you)
Mistakes - finishing combo on block or special on block
Effects(Taunt, Intimidate, Infuriate) - Classic Titania, Starky , gladiator....

how does light intercept , parry & Side step work ?
As the devs have often stated and the Deep Dive reiterates, the AI doesn't understand intercepts, or any tactic really.

But consider what an intercept is. How do we do intercept? We wait for the AI to come to us, and the moment our attack could actually reach the defender (depending on whether we are doing a dash forward or a light intercept) we throw our attack into that window of opportunity, hoping it lands before the AI can throw its own attack.

The AI doesn't understand "intercept." But it does understand "attack if the player decides to approach us and is within range." Since the AI is, as the Deep Dive mentions, extremely fast and precise, if the AI decides to hit you, it will throw an attack at you at the first instant such an attack would actually land. The AI doesn't have to hope, it calculates.

What does "throw an attack the moment the player enters into range of the attack" sound like? The AI just wants to hit you, but it won't throw a whiff. It will wait until the attack would actually land, and throwing an attack as the opponent approaches you to throw an attack themselves is what we call an intercept.

So the AI isn't trying to intercept. It is just trying to hit us quickly. But perfect timing on an attack while we are entering into range is indistinguishable from an intercept.

Same way AI doesn't try to punish them try to block that turns into parry 😐

PickL1e89

what change ai behavior :

Ai profile - ez for Arena , hard probably for rest
Champion Specific Tuning - chavez , mysterio, doom(heavy heavy)
Aggression, - chavez , mysterio (keep cornering you)
Mistakes - finishing combo on block or special on block
Effects(Taunt, Intimidate, Infuriate) - Classic Titania, Starky , gladiator....

how does light intercept , parry & Side step work ?
Venom and CMM are also ridiculous aggressive.

Unless a nodes say, all characters ai should have same level of aggression. Unless a node says so characters ai shouldn't follow favoured patterns, learn player actions or favour certain sp attacks

KabamDORK

Probably irrelevant but how does this kind of stuff sits in with the explanation given?

https://youtu.be/ubbniSiyzzs?feature=shared

I believe this will be better answered in the part 2 of the series. This falls in the camp of what the AI controls and has access to vs what decisions it can make which is what the post was largely about.

SummonerNR

Probably irrelevant but how does this kind of stuff sits in with the explanation given?

https://youtu.be/ubbniSiyzzs?feature=shared

I believe this will be better answered in the part 2 of the series. This falls in the camp of what the AI controls and has access to vs what decisions it can make which is what the post was largely about.

Even though you say more to come will be in “Part 2” of the AI discussion, I think at least part of it (the 2nd example in video, Ronan vs Spidey) may fall under what you had already mentioned last month in the initial discussions about Intercepts (which maybe was a “Pre-logue to AI, or Part Zero intro”).

https://forums.playcontestofchampions.com/en/discussion/391180/medium-intercept-issue#latest

In particular, if a Medium has to incorporate a Dash as part of the process, because you’re not near enough for the Medium hit itself, that it then involves both the Dash and Medium hit, which if I’m reading the earlier link below correctly, seemed to open up the possibility of a defensive counter right inbetween those (the Dash part and the Medium Hit part, of your single Medium swipe action)

Or more so if a single Dash is not enough to close the gap, you mentioned that it goes into a Loop of successive Dashes to close a large gap, and then perform the actual Medium hit. More time for defender to squeak in something ??

Not sure if that part would have already been resolved with the couple changes that were made then (the original fix, followed by fixing what then resulted in 60 fps follow up timing bug), or if the specific Ronan vs Spidey example in video would still occur now because of it being evade related vs Dash related ?

TotemCorruption

1. Very informative and transparent post by Kabam. I'm really enjoying the behind-the-scenes communication lately.
2. Even if the AI isn't conscious of many things, if certain things feel intentional to players, then the distinction only makes a difference to the coders.
3. The Serpent knows he's The Serpent....

benshb

It's also interesting how recently the AI started "special-intercepting", sometimes even backing up 1 or 2 steps, blocking, and when the player dashes in...boom. I'd say it's a pretty low-chance combination of random movement roll 🤷

Average_Desi

Probably irrelevant but how does this kind of stuff sits in with the explanation given?

https://youtu.be/ubbniSiyzzs?feature=shared

Is this still the case, I can't remember the last time I was punished for an evade unless I kept attacking .

Average_Desi

They could have given the "AI" a new name with 2.0. we could have killed all the "AI is learning and parrying me" posts .
....I kinda wanna push for this now xD
I vote Gary… imagine all the posts about how unfair Gary is or how Gary is intercepting too much lol

Gary is too stupid to learn. He's just the same as he was.

PickL1e89

The intercepting is off. The ones where they dash back, delay then come in are fine.
The harder intercept seem unlikely. Either everyone experiences lags or the ai is tuned to be too fast

ahmynuts

Can we have karatemike make these deep dives into videos with the deep dive voice

DRTO

They could have given the "AI" a new name with 2.0. we could have killed all the "AI is learning and parrying me" posts .
....I kinda wanna push for this now xD

Would it be copyright infringement if you call it JARVIS?

Frumpy_geezer

The AI doesn't understand "intercept." But it does understand "attack if the player decides to approach us and is within range." Since the AI is, as the Deep Dive mentions, extremely fast and precise, if the AI decides to hit you, it will throw an attack at you at the first instant such an attack would actually land. The AI doesn't have to hope, it calculates.

What does "throw an attack the moment the player enters into range of the attack" sound like? The AI just wants to hit you, but it won't throw a whiff. It will wait until the attack would actually land, and throwing an attack as the opponent approaches you to throw an attack themselves is what we call an intercept.

It's actually less "intelligent" than this. There's not an "and" that takes into account future actions. It's the current state of game in that exact frame. So, it's more like "the player IS approaching and MAY be in striking distance: attack (50%) or block (25%) or dash back (15%), etc" that's when you get intercepted or parried or whiff or land the attack.

JESUSCHRIST

It's obvious that different champions have different weighted behaviour tuned to each champion's play style

Like how Mordo fights differently from Jubilee (I can't think of a worse defender in the game)

JESUSCHRIST

Range awareness contribute to XL champions being more aggressive as they can reach you more often and more easily

I talked about XL champions having an advantage in game with their range

Like how Onslaught is super aggressive all the time

XL defenders FTW

JESUSCHRIST

Range awareness contribute to XL champions being more aggressive as they can reach you more often and more easily

I talked about XL champions having an advantage in game with their range

Like how Onslaught is super aggressive all the time

XL defenders FTW

Size does matter....

Elux

Seems like kabam are making excuses for the ai

DNA3000

The AI doesn't understand "intercept." But it does understand "attack if the player decides to approach us and is within range." Since the AI is, as the Deep Dive mentions, extremely fast and precise, if the AI decides to hit you, it will throw an attack at you at the first instant such an attack would actually land. The AI doesn't have to hope, it calculates.

What does "throw an attack the moment the player enters into range of the attack" sound like? The AI just wants to hit you, but it won't throw a whiff. It will wait until the attack would actually land, and throwing an attack as the opponent approaches you to throw an attack themselves is what we call an intercept.
It's actually less "intelligent" than this. There's not an "and" that takes into account future actions. It's the current state of game in that exact frame. So, it's more like "the player IS approaching and MAY be in striking distance: attack (50%) or block (25%) or dash back (15%), etc" that's when you get intercepted or parried or whiff or land the attack.

You're correct that as far as I am aware the AI does not predict the future, nor did I say "will be in range" I said "*is* in range."

However, as far as I know the AI does not guess either. There's no "may" be in striking distance. The AI doesn't see a dynamic game. It only sees each individual static frames of activity. You are either in range, or not in range (as the AI sees it). The AI has no understanding of "might be in range."

While the AI doesn't "predict" actions, it can be programmed to implicitly act on predictions because one of the things the AI is stated to understand is animation state. That means the AI can theoretically be programmed to act differently if the player is in the animation state of dashing forward or dashing backward. It is not predicting the motion, it is only looking at the state of the game at that instant of time. But the animation state implies certain things will happen in the future in some cases.

Herbal_Taxman

I am intrigued by what someone was describing above about a player’s dash across the screen. If our dash is of a certain length, it’s actually several consecutive actions even if we only swipe once.

IF this is correct (not sure it is), that would help explain the intercepting. For a frame perfect AI, consecutive actions making up what we believe to be a single input for dash would create several chances for the AI to guess correctly and attack. Thus performing what we perceive to be an intercept.

altavista

I am intrigued by what someone was describing above about a player’s dash across the screen. If our dash is of a certain length, it’s actually several consecutive actions even if we only swipe once.

IF this is correct (not sure it is), that would help explain the intercepting. For a frame perfect AI, consecutive actions making up what we believe to be a single input for dash would create several chances for the AI to guess correctly and attack. Thus performing what we perceive to be an intercept.

There is also the situation in the opposite direction, where I swipe back to dex (For example, I'm using Okoye, so want to dex so I can place an Ensnare). During my one swipe, my opponent seems to do this weird toggling between idle-block-idle-block-idle-block. So the AI is 'rolling' a bunch of actions while I seem to only be able to do my one dex action.

Avenger_A1

I saw a clever post. Seems it's getting the desired response. I wish I had more time to write about why so many things don't add up. Anyway, we'll see.

DNA3000

I am intrigued by what someone was describing above about a player’s dash across the screen. If our dash is of a certain length, it’s actually several consecutive actions even if we only swipe once.

IF this is correct (not sure it is), that would help explain the intercepting. For a frame perfect AI, consecutive actions making up what we believe to be a single input for dash would create several chances for the AI to guess correctly and attack. Thus performing what we perceive to be an intercept.
There is also the situation in the opposite direction, where I swipe back to dex (For example, I'm using Okoye, so want to dex so I can place an Ensnare). During my one swipe, my opponent seems to do this weird toggling between idle-block-idle-block-idle-block. So the AI is 'rolling' a bunch of actions while I seem to only be able to do my one dex action.

There's two different things going on here. There's the notion that some "actions" are actually multiple actions, and there's the notion that some actions are a single action that is implemented with a sequence of motions.

A single "action" the AI takes could be multiple inputs that blend smoothly into a single observed act. In theory, if a human was fast enough or smooth enough with their inputs they could also do this. But then there's also the case that most inputs that trigger actions do some sort of animation associated with that action, and most animations take place across many visual frames. Some animations are not interruptible, or not interruptible by certain other actions, so once you start them you're "locked in" to them. You can't change your mind in the middle of a medium attack to trigger a light attack, for example, but you can interrupt a medium to trigger a special.

Because the player can only act on certain animation frames if they are doing certain things, and the AI also can only act on certain frames if they are doing certain things. this creates a complex web of windows of time when certain things can occur, and that's what I believe Kabam was alluding to when discussing intercepts.

The "block seizures" the AI exhibits is probably just a visible side effect of the fact that blocking is an act that is relatively fluid: it doesn't lock you into block for very long, so the AI has the opportunity to decide to block and then not block, in rapid succession. My guess is the AI is in a state where one of the valid options is to block, so it does. This puts it into a state where there is really only two options: continue to hold block or release block (it cannot act while simultaneously holding block in the same instant, just like we can't). That state gives it a high probability of wanting to so something, so the odds of holding block are low and the odds of releasing block (to put it into a state where it can do something else) is very high, so it releases block. But that then puts it into a state where all the programmed options either can't be done or have a very low chance of occurring, so RNG quickly rolls hold block again. And you end up in a loop of hold block/release block/hold block. It can do this very quickly because blocking doesn't have very much time when you are locked in and can't act.