Fighter AI Realism vs Difficulty

[Kingslayer04]

Re: Fighter AI Realism vs Difficulty

The problem is that there isnt a venue to correct the perception. Not everyone goes to sites like OS. In fact most dont.

So lets say we didnt have this convo. You play UFC 4 and you see Pettis shooting for a takedown in every fight (he lands less than 1 per fight on average). Now your perception before this convo was that Pettis is mostly a kickboxer. So you see that he is attempting takedowns and you come to OS to make the devs aware. Skynet says "He actually averages over 1 attempt a fight". You realize he's right and you are fine with it.

Now lets say you dont go to OS.

The average player doesnt come here. So they will just think that the AI is screwy because of their perception of Pettis.

Same goes for guys like Rampage. He's averaged 1.5 takedowns per fight but people see him as a standup brawler so if he is regularly attempting takedowns, some will think something is wrong with the AI.

Its a balance and you have to be careful with it because of perception.

I do get your point, however why shouldn't we strive for realism? I infer from your points that you believe (correct me if I'm wrong) this adjustment to Pettis would be a change so significant that it will turn casual players off from the game. However, Garbrandt AI fights like a wrestleboxer, I brought that up and you referred me to his Briones fight, fair enough. But what about the target audience you mentioned, it's the exact situation you described, however in this case Garbrandt resorts to wrestling IRL far, far less than his AI does, whereas Pettis does that much more heavily IRL, which his AI doesn't reflect.

[aholbert32]

Re: Fighter AI Realism vs Difficulty

I do get your point, however why shouldn't we strive for realism? I infer from your points that you believe (correct me if I'm wrong) this adjustment to Pettis would be a change so significant that it will turn casual players off from the game. However, Garbrandt AI fights like a wrestleboxer, I brought that up and you referred me to his Briones fight, fair enough. But what about the target audience you mentioned, it's the exact situation you described, however in this case Garbrandt resorts to wrestling IRL far, far less than his AI does, whereas Pettis does that much more heavily IRL, which his AI doesn't reflect.

Every thing is a case by case basis. Lets take Cody. Cody was an HS and briefly a Div 2 wrestler. When the devs created his template it was before the Dillashaw 1 fight. At that time, he attempted no less than 3 takedowns in every fight that went past the 1st round (Briones, Brimage, Cruz). The only fights he didnt attempt were the first round KO run he went on.

So the perception is Cody is a striker with a wrestler base and who comes from a wrestler's camp (Alpha Male).

Take Pettis: Kickboxer by training. Most remember his standup fights and his subs. Many think of him as a poor wrestler. Most of his takedowns have come from the cage or defensive wrestling. Comes from a camp that focuses on striking.

So the perception is he isnt much of a wrestler.

Also there are game limitations that affect this too. Like I explained before, The AI doesnt take its opponent in consideration when gameplanning so how would the average person look at the AI if Pettis was attempting takedowns against Khabib?

[Kingslayer04]

Re: Fighter AI Realism vs Difficulty

Every thing is a case by case basis. Lets take Cody. Cody was an HS and briefly a Div 2 wrestler. When the devs created his template it was before the Dillashaw 1 fight. At that time, he attempted no less than 3 takedowns in every fight that went past the 1st round (Briones, Brimage, Cruz). The only fights he didnt attempt were the first round KO run he went on.

So the perception is Cody is a striker with a wrestler base and who comes from a wrestler's camp (Alpha Male).

Take Pettis: Kickboxer by training. Most remember his standup fights and his subs. Many think of him as a poor wrestler. Most of his takedowns have come from the cage or defensive wrestling. Comes from a camp that focuses on striking.

So the perception is he isnt much of a wrestler.

Also there are game limitations that affect this too. Like I explained before, The AI doesnt take its opponent in consideration when gameplanning so how would the average person look at the AI if Pettis was attempting takedowns against Khabib?

Oh yeah, about that last point - definitely. Besides, I'm not campaigning for Wrestler Pettis, we're just talking. I would disagree about the perception on Cody though. I think most people see him as an exciting boxer with fast hands, who mixes in some kicks and that's it. When a casual player/fan (that goes for me too) turns UFC 3 on and picks TJ or Cruz and wants to face Cody for a fun stand-up fight, what all 3 fights between Cody and those 2 were, he'll get a fight where Cody takes him down...and again...and again. And he'll clinch him too. And the round is over. To be fair that's my perception of him also. I did have a look at that Briones fight though, and I definitely saw him actively trying to grapple, which was new for me. I should see that damn Cruz fight again and see those 3 shots you are talking about. But yeah, I do think he needs a change in AI behaviour. 1 takedown in 3 rounds, 2, maximum 3 in 5 rounds sounds acceptable to me, but right now he's too much of a wrestler compared to how his fights actually go.

[Phillyboi207]

Re: Fighter AI Realism vs Difficulty

Thats a Skynet question. The reason this sticks out so much in my head is because of UFC 2. When we were testing UFC2, it used to annoy the **** out of me that Anderson Silva's AI would occasionally shoot for a takedown even against guys like Chael Sonnen. I asked the devs at the time could that be prevented. Could the AI just see the OVR or even things like TDD and just say "Not gonna attempt a takedown". The answer back then was no.

That could be part of the reason for developing individual fighter AI templates so stuff like that doesnt happen. Now that works for people like Pettis who rarely if ever attempt takedowns. It doesnt work for the fighters like GSP who are well rounded and whose strategy is based on the matchup.

A guy like Khabib is going to attempt a takedown against everyone but a guy like GSP isnt and that is a hole in the AI that I hope gets fixed if there is a next game.

Can we get a bit of an explanation for how the “goals” work for the AI. I’ve seen you mention it several times. Like how a fighter can become more or less aggressive if they’re losing the fight.

Is this something that is tuned per fighter/ template or does the AI share goal tendencies?

[aholbert32]

Re: Fighter AI Realism vs Difficulty

Can we get a bit of an explanation for how the “goals” work for the AI. I’ve seen you mention it several times. Like how a fighter can become more or less aggressive if they’re losing the fight.

Is this something that is tuned per fighter/ template or does the AI share goal tendencies?

Skynet can explain it WAY better than me. The goals are tuned by fighter and can be affected by whats going on in the fight. Some are tuned to be aggressive throughout the fight whether they are winning or losing. Some may be aggressive but change and become more conservative later in the fight and vice versa.

[Skynet]

Re: Fighter AI Realism vs Difficulty

On a side note, if the auto getups from knockdowns are removed, could the AI sometimes stay on the ground and wait for the referee to stand them up?

IRL, we see fighters that have been dropped sometimes just lay on their backs for a bit to clear their head.

Would be a good way to help fights go the distance.

Someone like Werdum should do this almost all the time.

I can neither confirm nor deny the existence of work pertaining the the AI's tendency to stay or not stay on their backs given certain fighting styles when put into certain questionable scenarios. I can definitely not confirm or deny that fighters prone to grappling and being on the ground will favor this tactic more so than others.

[Skynet]

Re: Fighter AI Realism vs Difficulty

Cool, thankyou.

Any chance you could recommend he perhaps have another look at the universal AI backfist rates?

I dunno about everyone else but either they're secretly OP or the AI is WAY too good at timing them.

They seem to know how to counter almost everything with them.

Advancing jabs, roundhouse kicks, teeps off the opposite side to the spin, hooks, low kicks (this one is totally acceptable of course) ect ect.

It's not that a backfist doesn't work in these contexts occasionally IRL, its just EVERYONE who has them uses them as the go-to counter for these moves.

This results in guys who have, but rarely, used the backfist will use it like Paul Felder.

I'm talking Masvidal, Arlovski, Manuwa, guys like that. Aswell as the lighter guys who actually do utilise them semi often.

Pettis and MM oddly rarely use them. Despite having them, dunno if it's because they have the slower lead spin backfist or if its their unique AIs.

Now I admit I just got KO'd in under a minute by Arlovski spin fisting me on two seperate high kick attempts with the 2nd flash KOing me, so its a slightly personal matter haha :P

But I can't be alone in thinking the backfist counter rates are still waaay to high almost universally can I?

Mainly in reference to Pro AI, the Hard AI throws them naked and not as counters pretty often. But Pro/Legendary use them as counters to a LOT of strikes.

Jimi Manuwa is a good example of the issue. He uses SBF counters 24/7 on pro.

I have actually noticed this, and it is rather odd, isn't it? I'm curious as to the cause, but as a first guess I would think the SBF has a rather short windup time with a relatively high interrupt value. Which means it is a viable strike to interrupt with in many cases.

The AI does indeed compute on pro/legendary what strikes will let it counter, when it wants to, and then randomly picks from them. Maybe the AI is pointing out some holes in the strike tuning, and the SBF is indeed stronger than intended :P

[Skynet]

Re: Fighter AI Realism vs Difficulty

Can we get a bit of an explanation for how the “goals” work for the AI. I’ve seen you mention it several times. Like how a fighter can become more or less aggressive if they’re losing the fight.

Is this something that is tuned per fighter/ template or does the AI share goal tendencies?

In AI design, there are many different techniques and ways of achieving the right AI for the job. In gaming in particular, there are maaany different things that can govern AI. One popular approach is called goal oriented action planning (GOAP), which is one of the backbones of UFC 3's fighting AI. I'll just call it the planner.

The planner has an internal representation of important game data known as state, a bunch of steps it can take to change the state, and a number of possible arbitrary goals to choose from. Steps have preconditions: what state the game must be in to start them; effects: how the state will change if the action is taken; and an action: what to actually do if chosen. Goals also have actions and preconditions, but no effects. Goals and steps can often overlap, since they all really map to actions. Sometimes you want to to that action specifically and sometimes it's just a stepping stone, such as takedowns and getups.

Some example goals in UFC would be throw a strike combo, recover stamaina, or attempt an armbar from fullmount.

Some example steps could be maintain distance from the opponent, close distance to the opponent, transition from position A->B, or perhaps get up.

The first part of this system is how goals are actually chosen. There are two main parts here: a static tendency driven by fighter customization, and a dymanic tendency driven by fight IQ (kind of...). The static portion is just that: a flat tuning value that every fighter has and it never changes during a fight. The dynamic portion changes every frame based on the game, and is specific to the goal itself. A good example is the goal to recover stamina. Obviously, some fighters are more prone to staying fresh than others (static), but the more tired any fighter gets the more likely they are to take a quick breather (dynamic). In our game this means the lower the AI's short term stamina becomes, the higher the chance their next goal choice is to recover. These tendencies combine and are then used as a weighted random when picking a new goal.

If we expand on the strike combo goal, it has preconditions such as both fighters are standing, being within range, and with sufficient stamina. If all those conditions are true when the goal is chosen, then it applies the associated action right away. In this case, asking the AI to throw a combo (with some filters). If those conditions were not true, then the planner will build a theoretical list of steps to get the game state into the right place (the plan!).

Let's say the AI is currently in sub guard. The steps would be something like getup->recover stamina (which moves them away)->close the distance. If all those steps are successfully taken, then the AI's state will indeed match the strike combo goal conditions, and the combo will be requested.

In UFC 3's case, our planner does indeed just plan, and never initiates any actions itself. The performing of actual actions is handled elsewhere, in what we call the procedural action system, which handles with more fidelity how actions should in fact be carried out on a more moment to moment basis. This other system is what actually sends inputs to the controller, etc.

So tl;dr: UFC 3 uses an AI technique known as goal oriented action planning (GOAP), and combines both static fighter tendencies with moment to moment fight IQ tendencies to drive goal choice.

PS: striking aggression going up/down in the planner as occasionally mentioned is based on the concept that if the AI thinks it is winning on the feet by a decent margin, it will push the pace in an effort to finish the opponent off. If they are losing by a fair margin they will try and increase their one-off striking to get lucky/points. If the fight is reasonably balanced, they won't change much in this regard.

[AeroZeppelin27]

Re: Fighter AI Realism vs Difficulty

I have actually noticed this, and it is rather odd, isn't it? I'm curious as to the cause, but as a first guess I would think the SBF has a rather short windup time with a relatively high interrupt value. Which means it is a viable strike to interrupt with in many cases.

The AI does indeed compute on pro/legendary what strikes will let it counter, when it wants to, and then randomly picks from them. Maybe the AI is pointing out some holes in the strike tuning, and the SBF is indeed stronger than intended :P

If honest I do kind of think its a touch overpowered now that I've seen about 20 different ways it can be used to counter things haha. It feels a bit too easy to counter an advancing jab with a SBF, for example.

Another issue with it is, and this may be unique to Manuwa, but I've had a few times where I've been able to get him stuck on a SBF loop.

Basically by throwing the rear teep to the body, he would try to SBF counter like he does for the high treo, but the pushback makes his SBF mix.

However sometimes he'll always use the SBF to counter the body teep which results in him spinning and you kicking him out of range.

It can be easily abused to make him gas, however, he won't always use the SBF in this context

[AeroZeppelin27]

Re: Fighter AI Realism vs Difficulty

Jessica Penne seems to have pure striker AI.

She will go for standups 95% of the time, doesn't shoot and only goes to Thai Plum when clinching.

I've had her attempt 1 submission during a TD on the cage. But I've had several striker AIs do this occasionally so I don't think it means much.

So, to update:
Joe Lauzon (Pretty sure Aholbert said he still had the striker AI after the last patch, haven't personally tested)
Jake Matthews
Jesscia Penne

All have pure striker AIs.

Joe should prooobably be a sub specialist or the Diaz type of Brawler who will engage on the ground but won't necessarily try to get it there often (but without the Diaz output)

Matthews should be balanced/wrestler who attempts subs and Penne if I'm not mistaken should be quite grappling heavy and sub friendly.

[sdpdude9]

Re: Fighter AI Realism vs Difficulty

Skynet, can you explain why I am seeing this type of behavior frequently from fighters are that are typically more passive fighters? Machida is being hyper aggressive from the get go.

https://youtu.be/dnp7nA1c5Uw

The aggressive types are even worse about this in my experience. This is on Pro difficulty with no sliders. The ai is often so aggressive they are gassed by round 3, or even halfway through round 2.

[Kingslayer04]

Re: Fighter AI Realism vs Difficulty

Skynet, can you explain why I am seeing this type of behavior frequently from fighters are that are typically more passive fighters? Machida is being hyper aggressive from the get go.

https://youtu.be/dnp7nA1c5Uw

The aggressive types are even worse about this in my experience. This is on Pro difficulty with no sliders. The ai is often so aggressive they are gassed by round 3, or even halfway through round 2.

I played vs Machida AI with Uriah Hall, +1 on Stamina tax for both me and AI, Legendary difficulty. It was a counterstriking behaviour, low output, no complaints whatsoever.

[aholbert32]

Re: Fighter AI Realism vs Difficulty

I played vs Machida AI with Uriah Hall, +1 on Stamina tax for both me and AI, Legendary difficulty. It was a counterstriking behaviour, low output, no complaints whatsoever.

Now thats on legendary where the output is higher than on pro. FYI, that output will be even lower with the new patch.

[Kingslayer04]

Re: Fighter AI Realism vs Difficulty

Now thats on legendary where the output is higher than on pro. FYI, that output will be even lower with the new patch.

About Legendary, I did mention it to you yesterday, I really don't feel like it's crazy unrealistic or anything like that. The only difference was Anderson Silva, who was a combo magician (which I have no problems with) who, however, pressed forward, which I thought he shouldn't be doing to the extent he did. Could you explain once again what are the differences between Pro and Legendary, are fighters supposed to fight "out of character", etc.?

[aholbert32]

Re: Fighter AI Realism vs Difficulty

About Legendary, I did mention it to you yesterday, I really don't feel like it's crazy unrealistic or anything like that. The only difference was Anderson Silva, who was a combo magician (which I have no problems with) who, however, pressed forward, which I thought he shouldn't be doing to the extent he did. Could you explain once again what are the differences between Pro and Legendary, are fighters supposed to fight "out of character", etc.?

The info I have on this may be a bit old and my focus has always been on pro so I may be a bit incorrect. My understanding is that the pace on Legendary is higher than it is on pro. The AI throws longer and more complex combos on Legendary. The difficulty is increased and they will be much better at defending strikes, denials on the ground, subs etc.

The longer and more complex combos (depending on moveset) can cause some fighters to appear to be fighting out of character. For example, in real life you rarely see long combos. 2-3 strikes are the norm. What you will see occasionally on Pro and more on Legendary is the AI throw a 3 strike combo....slight pause....and throw another 2 or three strike combo. That is rare in an actual fight unless someone is hurt.