Fighter AI Realism vs Difficulty

[aholbert32]

Re: Fighter AI Realism vs Difficulty

How do you determine which slider set would be used? Would there be a list of some kind that both players choose from? I don't know, sounds weird to me.

It wouldnt be for ranked. It would be for quick match and other online modes. Lets say you and I wanted to have a sim match and Philly uploaded a slider set, we would just both use his set.

[Kingslayer04]

Re: Fighter AI Realism vs Difficulty

@Skynet, I'm not sure if I've asked this question before but here goes: is there a way to make the AI capable of recognising patterns? My biggest issue with fighting the AI is that you can't really set traps in order to trick it into something or have to outsmart it, everything just feels so random. There's also its master countering ability too, as well as it biting on every feint. The entire mental aspect is non-existent. I know this may be very advanced and not so easy to implement but is it possible? It could be a separate Fight IQ stat or something like that, where the higher the stat is, the more the AI will be able to make these reads faster than those with a lower stat - everyone should be able to pick up on a repeated combo eventually.

[Skynet]

Re: Fighter AI Realism vs Difficulty

@Skynet, I'm not sure if I've asked this question before but here goes: is there a way to make the AI capable of recognising patterns? My biggest issue with fighting the AI is that you can't really set traps in order to trick it into something or have to outsmart it, everything just feels so random. There's also its master countering ability too, as well as it biting on every feint. The entire mental aspect is non-existent. I know this may be very advanced and not so easy to implement but is it possible? It could be a separate Fight IQ stat or something like that, where the higher the stat is, the more the AI will be able to make these reads faster than those with a lower stat - everyone should be able to pick up on a repeated combo eventually.

Hey mate. While it's certainly possible it's by no means trivial to recognize combo patterns, and the consequences of doing it aren't always as shiny as you expect. What you're really asking for is that the AI become good at predicting and countering repeated combos over a long period of time (multiple minutes/rounds).

At the moment, the AI has the ability to recognize single strike/target patterns in the short term, meaning if you throw 5 kicks to the body, they'll expect the 6th kick to be body as well, etc. This pattern recognition behavior is only split by punch/kick and head/body. We actually had it more granular than this, and the AI started feeling too robotic and cheesy. Effective prediction isn't always a good thing...

Now, that's got nothing to do with actual combo prediction. If 90% of the time you throw a 1-2 you follow up with a 3, then humans will indeed learn that and counter it. Building a system to recognize, and duly respond, to such predictions is not something I've put a lot of time into simply because the scope of that effort is quite substantial, and the benefits haven't been worth the effort. Especially if it ends up making the AI feel more robotic than human.

That being said, it's one of the higher skill tactics that I'd love to build into the AI at some point and experiment with, because I think it has the potential to truly transform the AI experience for moderate to skilled users. Fingers crossed the opportunity arises in the future

[Kingslayer04]

Re: Fighter AI Realism vs Difficulty

Hey mate. While it's certainly possible it's by no means trivial to recognize combo patterns, and the consequences of doing it aren't always as shiny as you expect. What you're really asking for is that the AI become good at predicting and countering repeated combos over a long period of time (multiple minutes/rounds).

At the moment, the AI has the ability to recognize single strike/target patterns in the short term, meaning if you throw 5 kicks to the body, they'll expect the 6th kick to be body as well, etc. This pattern recognition behavior is only split by punch/kick and head/body. We actually had it more granular than this, and the AI started feeling too robotic and cheesy. Effective prediction isn't always a good thing...

Now, that's got nothing to do with actual combo prediction. If 90% of the time you throw a 1-2 you follow up with a 3, then humans will indeed learn that and counter it. Building a system to recognize, and duly respond, to such predictions is not something I've put a lot of time into simply because the scope of that effort is quite substantial, and the benefits haven't been worth the effort. Especially if it ends up making the AI feel more robotic than human.

That being side, it's one of the higher skill tactics that I'd love to build into the AI at some point and experiment with, because I think it has the potential to truly transform the AI experience for moderate to skilled users. Fingers crossed the opportunity arises in the future

Thanks for the reply - hopefully it does!

Edit: what I generally want is for there to be rhyme or reason as to why certain things happen, cause and effect, some level of Fight IQ being employed, from both myself and the AI. Perhaps this can be achieved to some extent through different means, not necessarily a pattern recognition system out of thin air. But yeah, hopefully you guys can come up with something in the way of that.

[Kingslayer04]

Re: Fighter AI Realism vs Difficulty

Another one for @Skynet

We have talked about the AI being a countering god, but there's another side to the coin too: it's also extremely easy to trick. As soon as you get in punching range it'll start swinging. You can then do anything you like until its combinations are over: slip, back away, block, look to counter. As a result, it is extremely easy to gas it in a round or two. It also heavily bites on any and all feints which leads to the same result, arguably even faster. I constantly switch between Pro and Legendary to find the more challenging option but I usually find that on Pro the AI is often too static (tricking it into throwing is still there though), and on Legendary it gasses even faster due to throwing at air even more.

Also, grappling oriented AIs constantly go for takedowns without trying to set them up in any way, to pace themselves or to look for openings. As a result I get a gassed Frankie Edgar midway through Round 2 and when the fight ends he's 4/23. When they do get the TD/Clinch they always look for the most dominant position available: Crucifix from SC or Thai Clinch. They then proceed to attempt it ad hominem no matter how many times you deny it in a row. It's all just pretty mindless.

Hopefully these are addressed in the next game. Cheers!

[johnmangala]

Re: Fighter AI Realism vs Difficulty

I'm interested in seeing an unbeatable AI difficulty, if possible.

[Skynet]

Re: Fighter AI Realism vs Difficulty

I'm interested in seeing an unbeatable AI difficulty, if possible.

It's mildly interesting, but honestly not fun. In fact, it quite a frustrating thing. They just go god mode and start perfectly countering stuff. It's not a competition, and doesn't feel like a fair fight at all.

It's quite immersion breaking.

[WarMMA]

Re: Fighter AI Realism vs Difficulty

I'm interested in seeing an unbeatable AI difficulty, if possible.

Yh like Skynet is saying, it won't be fun. I haven't seen it obviously, but I can just imagine it will just be a counter machine and counter almost 100% of your offense. It will do the perfect counter for everything you do maybe almost 100% of the time. So i'm sure it won't be fun and it definitely won't be realistic.

[johnmangala]

Re: Fighter AI Realism vs Difficulty

It's mildly interesting, but honestly not fun. In fact, it quite a frustrating thing. They just go god mode and start perfectly countering stuff. It's not a competition, and doesn't feel like a fair fight at all.

It's quite immersion breaking.

Yh like Skynet is saying, it won't be fun. I haven't seen it obviously, but I can just imagine it will just be a counter machine and counter almost 100% of your offense. It will do the perfect counter for everything you do maybe almost 100% of the time. So i'm sure it won't be fun and it definitely won't be realistic.

Fun is objectively subjective. It may be true that many people haven't been classically conditioned, but there are still some people that enjoy that kind of challenge.

Not exactly referring to the god mode like countering ability the AI demonstrates. To me it feels like the AI is restrained too much by templates... I am referring to fighting an AI that's still vulnerable to being hurt in similar ways but just keeps surviving- like a gauntlet.

The issue imo is the focus on perfect counters, it's been immersion breaking (unrealistic) for many for many years. Instead maybe more sliders which tune such things to the gamer's preference. There are already presets which are defense only (block, sway, move)- balancing these tuner sets can provide a variety of combinations. I feel an unbeatable AI can be more immersive depending on how it's tuned.. imagine fighting Prime Fedor/Jon Jones as the can.

It's true you'd eventually lose but it'd be like fighting the AI unbound... no template but it IQs you then executes => God mode.

[Solid_Altair]

Re: Fighter AI Realism vs Difficulty

- Countering should remain the same but the AIs unique combo selection should always favour the most likely to land strikes on the hardest difficulty, and only throw a wild shot like a spin jump kick when there is no counter opportunity for the player. Too many times I can catch legendary AI as they throw a silly strike by using fast strikes like the jav.

Great feedback.

That part surprised me a bit. Do you really think that happens too often?

[johnmangala]

Re: Fighter AI Realism vs Difficulty

Great feedback.

That part surprised me a bit. Do you really think that happens too often?

Any reason in particular you blatantly ignored my feedback again?

This is a pattern with y'all, so many of my ideas are in the game but y'all act like yall soley came up with it.

Example: Martial Mind liking my posts like that Hernandez KO over Dariush, Silva moves, etc and adding those things..

Obviously these ideas aren't patented and ideas can overlap. But then to turn around and have guys like you and aholbert say y'all never have looked at my posts and said that's a 'good idea'.. that they only "inspire" the devs.

[aholbert32]

Re: Fighter AI Realism vs Difficulty

Any reason in particular you blatantly ignored my feedback again?

This is a pattern with y'all, so many of my ideas are in the game but y'all act like yall soley came up with it.

Example: Martial Mind liking my posts like that Hernandez KO over Dariush, Silva moves, etc and adding those things..

Obviously these ideas aren't patented and ideas can overlap. But then to turn around and have guys like you and aholbert say y'all never have looked at my posts and said that's a 'good idea'.. that they only "inspire" the devs.

Why does he have to respond to your feedback?

For example, I think an unbeatable AI is an unbelievably bad idea. I think it would be a waste of resources. Now typically, I would address that in the forums and I would have a back and forth about it.

What did I do with you? I saw it and didnt respond.

Wasnt in the mood to argue about it and saw that Skynet responded so I moved on.

Not to speak for Solid but Aydin's feedback was better. He has consistently provided AI feedback since release. He knows his ****. I liked his post better. Maybe Solid did too.

Finally, you dont know the intent of a "like". I've liked posts that are completely new ideas that no one has thought of. I've liked posts that support things I've told the devs I want in the game. I've like posts that I know wont be added to a future version of the game. I've liked posts sometimes as a clue that this is something that could be in the game.

Just because Martial likes a post of yours doesnt mean that the devs added it because they saw your post or he took that idea back to the devs. It just means he liked your post.

Anyway, lets not derail this thread. Its one of the few threads where a dev gives direct feedback regularly. You got a response from Skynet...be happy with that.

[Kingslayer04]

Re: Fighter AI Realism vs Difficulty

I think the problem with it is IRL you tend to get a feel for the opponent first before attempting something like that, strikes like the handplant roundhouse, naked rear overhand, spinning attacks, even naked head kicks.

Here I can hold off for a while, peppering with simple strikes and just wait, sooner or later they will throw something big with a large counter window and I can catch them.

Whereas in real life, you'd be highly unlikely to attempt a move like say a handstand roundhouse, here if a fighter has it in their arsenal, sooner or later they will throw it regardless of how badly they were performing with the basic strikes.

Some fighters do just throw it out of the blue IRL, but the risk reward is quite high, and ingame it's easy to catch them once you have learned all the animation cues.

One example is the overhand. I find myself consistently able to land a head kick as the AI throws an overhand just by distance management. The AI might not recognise the trap I'm setting, whereas a player would.

IRL, if you throw an overhand and don't find yourself getting anywhere near the mark, you either keep spamming and get hurt, make adjustments to close the gap or better disguise the shot or abandon it for a different strike.

Here afaik, it's difficult for the AI to recognise to make an adjustment. Like how can they really figure out what the player is up to? I guess that's the limitation to it right. Then it's perhaps about creating complex patterns for the AI to use for example: jab overhand > player retreated on the jab so the overhand missed > next time the AI will throw a double jab overhand to cover the distance > player does something else to avoid the overhand so the AI does a strike cancel as it sees the shot missing

Perhaps if they are able to recognise the defensive patterns against the more basic shots, then they could throw in a wild one later down the line, using strike cancelling whenever a big shot is likely to miss.

Perhaps just tuning the AI to know the most efficient methods of combos and striking patterns, especially on the highest difficult and sticking to those rather than throwing what we'd see the dude do IRL. But then we lose that fighter authenticity unfortunately unless they are able to closer tune the meta to match real life.

For example Stephen Thompson IRL has his knee chambered constantly at a long range, so his side kick, back kick, front head kick, side oblique kick, comes out at a flash and has very specific movement tailored to it and functionally sound. Whereas ingame it visually looks good, functionally it doesn't work the same. So that's something to really look at for future if possible. I mean you would actually need an expert to really explain why he is so effective although some dudes here could probably explain it too.

It's like programming chess moves. We need to find better patterns for the AI to use on higher levels and work around what the player is doing and adjust based on what their AI template is capable of rather than giving them matrix timing or 100% block success.

Perhaps with more gameplay depth, things will get more complicated and more difficult. I mean, I'm probably a minority in thinking that blocking is too easy. But something like the Gaethje block front and side, up and down being separate would really make the game a lot harder vs the AI and far more interesting!

I would love for pattern recognition to be a thing, I made a post about it some time ago too. It would also be great if the AI thought about its stamina and recognised when it's a good idea to slow down or clinch to conserve it.

Another thing that's absolutely huge is the AI recognising which fighter it's up against. It shouldn't be looking to take down elite grapplers unless it is an elite grappler itself. Also, even if it isn't up against someone you don't want to be on the ground with, it should recognise when it is unsuccessful with that tactic (or any other), instead of attempting mindless takedowns and clinch entries until it's absolutely gassed. I hope that's worked on.

[Solid_Altair]

Re: Fighter AI Realism vs Difficulty

I think the problem with it is IRL you tend to get a feel for the opponent first before attempting something like that, strikes like the handplant roundhouse, naked rear overhand, spinning attacks, even naked head kicks.

Here I can hold off for a while, peppering with simple strikes and just wait, sooner or later they will throw something big with a large counter window and I can catch them.

Whereas in real life, you'd be highly unlikely to attempt a move like say a handstand roundhouse, here if a fighter has it in their arsenal, sooner or later they will throw it regardless of how badly they were performing with the basic strikes.

Some fighters do just throw it out of the blue IRL, but the risk reward is quite high, and ingame it's easy to catch them once you have learned all the animation cues.

One example is the overhand. I find myself consistently able to land a head kick as the AI throws an overhand just by distance management. The AI might not recognise the trap I'm setting, whereas a player would.

IRL, if you throw an overhand and don't find yourself getting anywhere near the mark, you either keep spamming and get hurt, make adjustments to close the gap or better disguise the shot or abandon it for a different strike.

Here afaik, it's difficult for the AI to recognise to make an adjustment. Like how can they really figure out what the player is up to? I guess that's the limitation to it right. Then it's perhaps about creating complex patterns for the AI to use for example: jab overhand > player retreated on the jab so the overhand missed > next time the AI will throw a double jab overhand to cover the distance > player does something else to avoid the overhand so the AI does a strike cancel as it sees the shot missing

Perhaps if they are able to recognise the defensive patterns against the more basic shots, then they could throw in a wild one later down the line, using strike cancelling whenever a big shot is likely to miss.

Perhaps just tuning the AI to know the most efficient methods of combos and striking patterns, especially on the highest difficult and sticking to those rather than throwing what we'd see the dude do IRL. But then we lose that fighter authenticity unfortunately unless they are able to closer tune the meta to match real life.

For example Stephen Thompson IRL has his knee chambered constantly at a long range, so his side kick, back kick, front head kick, side oblique kick, comes out at a flash and has very specific movement tailored to it and functionally sound. Whereas ingame it visually looks good, functionally it doesn't work the same. So that's something to really look at for future if possible. I mean you would actually need an expert to really explain why he is so effective although some dudes here could probably explain it too.

It's like programming chess moves. We need to find better patterns for the AI to use on higher levels and work around what the player is doing and adjust based on what their AI template is capable of rather than giving them matrix timing or 100% block success.

Perhaps with more gameplay depth, things will get more complicated and more difficult. I mean, I'm probably a minority in thinking that blocking is too easy. But something like the Gaethje block front and side, up and down being separate would really make the game a lot harder vs the AI and far more interesting!

Thanks for the detailed reply.

Would you say it's more about the fancy moves being easy to interrupt, or easy to evade and whiff punish?

[johnmangala]

Re: Fighter AI Realism vs Difficulty

Thanks for the detailed reply.

Would you say it's more about the fancy moves being easy to interrupt, or easy to evade and whiff punish?

Why does he have to respond to your feedback?

For example, I think an unbeatable AI is an unbelievably bad idea. I think it would be a waste of resources. Now typically, I would address that in the forums and I would have a back and forth about it.

What did I do with you? I saw it and didnt respond.

Wasnt in the mood to argue about it and saw that Skynet responded so I moved on.

Not to speak for Solid but Aydin's feedback was better. He has consistently provided AI feedback since release. He knows his ****. I liked his post better. Maybe Solid did too.

Finally, you dont know the intent of a "like". I've liked posts that are completely new ideas that no one has thought of. I've liked posts that support things I've told the devs I want in the game. I've like posts that I know wont be added to a future version of the game. I've liked posts sometimes as a clue that this is something that could be in the game.

Just because Martial likes a post of yours doesnt mean that the devs added it because they saw your post or he took that idea back to the devs. It just means he liked your post.

Anyway, lets not derail this thread. Its one of the few threads where a dev gives direct feedback regularly. You got a response from Skynet...be happy with that.

It's not about anyone responding in particular. It's about the general undermining of feedback that goes against the established meta. Like I said, there should be a division of game-samers. Y'all smear ideas of people like me then beg the question by not supporting your premise that an unbeatable AI would be a waste of resources. Y?

The thing is I am not in the ever in the mood to defend evolving this game. Y'all constantly force people like me to argue by smearing and running away. I have generally present ideas only for established actors to swoop in and defend their status quo.


It's proved that these ideas can lead to more $ as they accommodate more demand and meets basic needs.


Also you missed the point on devs taking credit for unpatented ideas people presented here. You acknowledged people like Romero, etc having contributed to this game in that same thread by the same criteria... In that same and similar threads devs (including GPD, MM, you, Zombie, Altair, Zhunter, etc) under the same criteria ideas I campaigned are already in the game too... for example single collar stuns... Y'alls echo chamber is showing.

I know my feedback is better since I've inspired this collective evolution. It already exists in the game, like many entrance animations and anonymous mode (UT). It requires some abstraction but the idea is rooted -> practical and novel.

I understand I have been tone deaf and my framing has sucked but I've been reflecting back what I see. I am trying to filter it better now.

An unbeatable AI

Spoiler

I am referring to fighting an AI that's still vulnerable to being hurt in similar ways but just keeps surviving- like a gauntlet.

The issue imo is the focus on perfect counters, it's been immersion breaking (unrealistic) for many for many years. Instead maybe more sliders which tune such things to the gamer's preference. There are already presets which are defense only (block, sway, move)- balancing these tuner sets can provide a variety of combinations. I feel an unbeatable AI can be more immersive depending on how it's tuned.. imagine fighting Prime Fedor/Jon Jones as the can.

It's true you'd eventually lose but it'd be like fighting the AI unbound... no template but it IQs you then executes => God mode.