Hi guys, I saw a lot of comments being really happy that I'm "back". I really appreciate that, but I want to be clear I'm not coming back to making consistent videos. While I'm nowhere near the scale of TH-camrs like Tom Scott, I felt similarly burnt out trying to make regular videos. From now on, I'm simply going to make a video when I get an idea I'm motivated enough to work on, keeping TH-cam as a hobby instead of as a side hustle. Thank you guys for the love on this video, I'm very excited for Tekken 8 so you should see more of me soon ❤
My ghost is stupid because it stops adapting at a certain point, I've started rounds almost exclusively with down-back 3 as Jin and my ai hasn't done it once. Among other things
regarding the overwriting of old data with new data: Eddy Wang from Damagexcaling tested and confirmed that the ghost can only learn a finite amount of data, after which it will not learn anything new he found that in order to learn new data, the old data must be deleted to make room for more. we're all hoping that the full release will resolve this issue by allowing old ghost data to be overwritten, but that does not appear to have been implemented currently
It makes sense to limit it in the demo. They could really do any number of things with the ghost feature on release and in the future. Really looking to see how this feature..... evolves.
That actually makes sense. I spent most of my time teaching my ghost giga shitpost tactics. But then I couldn't get it to learn how to run away at low health, get a large space, ki charge into rage art when they run ip to me. I guess I filled it's data already by that point.
He also pointed out how the ghosts have their own seperate rank display. And how they're all locked to 'beginner' in the demo. Maybe the limit of how much data is stored increases with rank
In Tekken 6 after every time you quit to the main menu offline it will ask you if you want to overwrite the ghost data.... So I'm sure T8 will have it. T6 ghost was/is very good. So if it's better than T6 ghost we will be very good. T6 ghost will do sidestep counter hit setups and full combos with wall carry to wall with wall combo. So yea....,.
During the LCQ, one of the comments in the stream said "I don't know who to root for here so I'm going for Shirdel since I like his videos". That made me find your channel, and man this video was amazing. Love the analytical style, production value, and entertainment factor. Looking forward to more deep dive content like this for the full release!
My hypothesis for the part in 9:05 (Since I studied Machine Learning and deep learning) is that there are weights (a random value) that determines the importance of the current node. The nodes with higher value are most likely kept and not discarded. Also another observation of mine, I played a lot against my Kazuya ghost without using DF3,2,1 string.. So the ghost never used it before, but when I try to do the mind game by holding the last input of that string to get plus frames, every single time my ghost punishes it perfectly with 112 and holds his block if I don't hold (Every single time) (Even though I didn't explicitly teach him to do that by doing it myself before) So I believe the ghost learns the frames of the moves, so it will punish any -14 move with the same punish you used for any other -14 move. To add on, I also believe the ghost learns combos near the wall (Although it took a long time before the ghost started doing wall combos correctly). Also, The ghost doesn't learn the sequence if the sequence ends the round, It took quite some time to purposefully use rage art in situations that wouldn't end the round in order for the ghost to learn to use rage art.
That's really interesting stuff! The last part about not learning if the sequence ends the round makes a lot of sense, that seems like a huge oversight to me. Regarding df321 mind-games, that could be related to the different states. Because the charged move would count as a different state, the AI might be immediately recognising that and punish you. See what happens if you don't go for the last hit at all.
There's a GDC talk "Designing AI for Killer Instinct" which describes the similar Shadow AI in that game. Their system matches your assertions perfectly - it performs no extrapolation. What they do is copy inputs from a replay that best matches the current situation. The situation is measured with hand-chosen metrics, and the best match is picked with a nearest neighbor algorithm with some randomness.
Finding analytical, very well produced videos like this for fighting games in general is extremely rare, and I'm very happy you're filling that gap. Keep up the phenomenal work!
I have a feeling that, if similar, new data is preferred over old data, simply because this mimics most players getting better as they play more. It'd be interesting to do a multi-hour test where you play to your best, folllowed by playing distinctly worse and seeing if it prefers new data. Or we ask namdy bamdy.
it never ceases to amaze me how developers just refuse to improve enemy AI. even in the one genre where the bots are literally meant to mimic human opponents and you only have 1 entity to work with, they just don't. if you're lucky they'll read your inputs and call it a day e.e
This is definitely a huge step in the right direction but I absolutely agree and have always complained about it. Hopefully the ai in tekken 8 isn’t horrible
The ghost feature has been incredibly interesting to me, i had a feeling it worked like you explained but didn’t have the knowledge to put it into words, really loved this deep dive and i hope to see the ghost feature further developed over time
Do you think the ghost data file size limit of 8321 KB is preventing it from gaining complexity? I noticed that the ghosts of top ranked players weren't very complex and were easily defeated by hyper-aggressive moves, and I saw that the downloaded ghost files are all 8321 KB. Could this be an artificial limit placed by the developers, or is it developed such that ghosts only have 8321 KB to work with? I've only tried a little bit of machine learning, but I found that neural networks for even simple games like Breakout or Pong could have larger file sizes than Tekken's, so I wonder what could happen if it had more space.
I'm not sure how accurate all of this is, because I've been fighting my own ghost lately, and I swear that the ghost uses some moves that I have never, ever used myself, or if I have, they have been accidental, so they must have some form of agency.
this is still infinitely better and much more humanlike/entertaining than input-reading & makes me wanna play tekken single player modes for sure actual learning takes human hand crafting to build in algorithms & on how to notice data patterns & how to respond to it. That requires a different type of coding. based on a very strong player is to have a bunch of matches. that'll be very valuable.
I'm a reinforcement learning specialist. I am quite intrigued by the Q-learning approaches of super ghost battle. If you wan we can have a talk about possible technical details of the implementation.
This AI sounds pretty entry-level to me. Which, seeing how this mode is an extension of Arcade Quest, makes sense. An introduction to the serious play style of characters for casual players that are easier to digest, as an online player that's stomping them out can easily switch to mix ups before they can adapt. I can dig the idea, because the bridge from casual to competitive still doesn't have the clearest path. I just wish it didn't remember sub optimal punishes so harshly, because most times I'll do a jab punish for a mix-up, or, honestly, laziness.
A point i want to share regarding question one, the ghost doesn't choose randomly but will prefer your optimal combo considering space between you and him and you and the wall, i saw that depending on how much i was away form the wall he preferred a combo that putted me on the wall instead of the most damaging without wall carry. Another thing that blew my mind is after watching my friend and trying it my self woth kazuya, also the ai consider how much health it has left, if a player hal low health and because of that gets more nervous dropping more combos, the ghost will drop the combos as well when at low health or stressfull situation, it would also do some kind of preference in okizeme or standsr situations as well as reflexes in matchups you never did, it was pretty rough when we tested it but it tried to do mimic to an extend, we only tested it for like 30 hours not even being that good in tekken, pretty little time to see how much potential it has
I main Yoshimitsu, he's been my main since 1995. Even though I have spent a good amount of time training my ghost, and using every possible move that yoshi has, the AI can't seem to learn past a certain point. I can do a few difficult combos, repeat it often, and it just won't learn it. But it's okay, it's still an interesting concept that I'm sure will take time to get better on the developers side.
Take a look at the video from GDC “Designing Shadow AI for killer instinct” - its absolutely brilliant and I’ve been theorycrafting on my own for ways of implementing this in a smash-like game. The Shadow AI from that game has the same idea as the ghost AI from this game
spent a few hours last night and today and it started learning my new mistakes and combos lmao so looks like it does overwrite some but also remembers some old combos
I kicked the shit out of my ghost over and over, mainly because I'm terrible on defense, but the one time my ghost had enough in the tank to throw it all back and win......it felt like I was playing against a little brother who finally picked up what I was doing rather than myself. But that was still worth a lot to me as I immidiately hit Rematch.
I see this mode as a leaning tool for people who hate to go into replay and rewatch their games. Its a fun mirror to see what you yourself are doing. For this purpose its very good.
Man I hope the next update allows your ghost to fight another one of your ghosts, like imagine a pro ghost vs another pro ghost. I honestly dont like the cpu thingy as the actions are too slow, ghost vs ghost is quicker and much more better (hope they add it)
10:22 THIS. This is the bane of my existence. I'm playing Jin not even day 2, I do Electric Thrusting Uppercut but whiff the EWHF and do df2 instead and the ghost will NOT let me live it down. It barely does the correct combo no matter how many times I do it.
Its a shame the ghost can only effectively learn mirror matchups. But I also laugh whenever my kazuya perfectly duck launch any 1,2,4,3 attempts I accidentally throw up.
This video was very intriguing to watch and it made me think of Super Smash Bros Ultimate's Amiibo training. It's rather similar to the Ghost and I think it might be interesting to look into on your own time as it varies enough just from the differences in games alone.
@shirdel the Killer Instinct reboot from 2013 (its on steam) has something called Shadow Lab that allows you to create an A.I. based on your replay data. these Shadow A.I. are just as advanced as Tekken 8's Ghosts, if not more advanced (they will literally learn to teabag), and in fact theres an hourlong lecture from GDC on youtube from a former dev of Killer Instinct about the creation of Shadow Lab mode!
So, I pretty much only play old tekken games. Which means that my fundamentals are really solid, but a lot of the new tactics based around air juggles and wall carry and wall play I never really learned, and am getting the hang of now. It really feels as though my ghost grasped the importance of these secondary tactics right around the same time I did. It overwrote old habits very quickly, and continues to do so. I would say that this either means the routing is very good, or it can start calculating what decisions lead to the overall highest damage output.
I like how one time when I played with ghost I used a rage art in a combo awhile back. Then later when ghost launched me it was in rage and did the same combo ending in rage art.
I’m a Tekken Vet & I’m really good at game 3 mains in red & purple ranks I love the Ai because it forces you to beat yourself thus making your main character better in execution & defensively
To me a HUGE issue is that you have to fight your ghost for that. It would be better if it learned as we play, but so far, you need to go to a specific part of the game for it to train.
One obvious flaw I found in the demo was it didn't properly learn to duck a high string. At the same time I was amazed by their small SS whiff punish. I was like no way they learned that from me 😂
If machine learning is used. My guess is as you play the game it makes predictions of how you'd act while you are currently playing. If it's correct it's rewarded.
Ghost has already crossed a threshold where it plays "player-like" which is enough for me. Sometimes I can't tell if I'm playing a ghost or a player. I knew right away that it wasn't a very advanced AI, but it's worlds different from the old type of dumb-AI that their CPU used to use, where the most it would do is 1 random move at a time. We're entering a new era of AI opponents in game design and I'm honestly quite excited for it. I can't wait for someone to make the next soulslike (hate that term but for the sake of brevity) type game except the bosses fight almost like PVP players would.
You gotta remember, they would want the AI to optimise its options as much as the individual does, not to play using your techniques but without your level of thoughtfulness. It's true that they have failed to implement intent, but they would want to copy the players intent, not optimal intent. If I press on minus, my ghost should do so too
OK, but important question. Will it know your inputs after you win a round to sayyy ki charge, Miguel's f3+4, or Raven's labyrinth over and over again? This is vitally important.
I really like this feature. While I understand frames and punish windows, I rarely use practice mode to learn to beat a certain characters, it gets boring. I rather play with or against the character, observe and adapt. With this system I can learn a character and when I pick a new one I can practice by playing against my AI of the previous one to skyrocket my overall knowledge of the whole cast by repeating this process
I'm not sure precisely what's going on, but I have noticed the ghost will occasionally do things I have never done. For example, I was playing Kazuya and did not one wavedash, EWGF or backdash cancel and the ghost was occasionally pulling them out.
Just started messing with mine. Progress: 1st match, cake 5th match, started using my combos 10th match, learned heat & rage 20th match, used moves I didnt know😅 30th match, blocked all i threw at it, theee in some grabs 40th match, counters my grabs 🥺 57th match, at this point its a fight to see who moves 1st...Ive played as Jin for yrs so I know how he fights. My only saving grace at this point is i target the frames where he cant block😂 i fear whats to happen next. I stopped teaching it
I'm no AI expert, but I'm a regular software engineer ...and my intuition (which might be wrong) is telling me a few things. 1. The file is probably always the same size because it's just a big list of weighted outputs for every predefined input. New data doesn't get "added" in the traditional sense -you're just changing numbers. 2. I think it's only using a certain amount of your most recent matches, as to reflect your current playstyle more accurately. 3. There's probably some basic AI agency, but only a Very Easy bot or something that things are layered on top of. I have a feeling it would need way more than a couple of matches worth of data to be able to play deterministically.
So far I'm a quite impressed on their offense but their defense leaves something to be desired. But I suppose much of that has to do with many players going full offensive and rarely blocking, especially against the CPU. There's no true replacement for playing against real human players but the ghosts are still more fun to fight against than the default AI.
I have hypothesis about how ghost ai dumps data and this will also solve why game data is not expanding expotentially. "If player forexample makes the best combo punisher 0% and 25% of time he makes some combo punish." Here I predict ai would make 0% of time the best combo punisher . Then he plays more and gives more info to a.i. In the next example. "If player forexample makes best combo punisher 25% and 50% of the time make some combo punish." Here I predict ai would make best combo punisher 100% of the time or maybe there is some variance adjusted? In time this results and correct button presses are like nodes that get activated and the most sophisticated ai ghost will never make input mistake. In this model of ai. The final best ghost could same amount of info or close to it whether the switches are on or off. So you save hardrive space. I am currently only advancing to test this ai but I am in terested in machine learning. Thank you for your time! Have a nice day.
could it be that there's a finite number of fighting options and your "ghost" is scored on all of those options. (Given a discrete scenario you throw sometimes. So, you establish a 4 out of 10 probability of throwing in that scenario. That "throw score" of 4/10 would mean that sometimes you throw... so when the ghost is in that situation, it throws 4/10 of the times ) The hard part about building ghost intelligence like that is identifying the scenarios and scoring the action. Refinements in that system could be thought of as "advancements in the intelligence".
So, basically the ghost is just a set of values for pre-determined (yet undisclosed) parameters? Every check it does it asks itself a question "what would be the best probable move to do after, given the current state", which it gets directly from the current state of the game mode (ue thing). I think that THATS THE EXACT RESON why it feel like your ghost does some "CPU"- level stupid stuff sometimes, especially when you do nothing, leaving it suddenly with little gamestate changes. My guess is that in the absence or in cases of low projected probability, it just resorts to its next move being "the default" for conventional decision-tree based bot. Sprinkle some salt (random-ish variation) and boom! I also think you're giving bamco more credit than they deserve when it comes to this thing. I highly doubt they've engineered a custom-tailored neural network just for the purpose of FG ai. They've probably mapped out all the moves (because they have to do it anyways lol) and states of the game (time, position, juggle, heat amt etc.), and used some pre-existing network to "teach" it with the game data as well as made it being able to execute stuff as a cpu-controlled pawn. In the demo, it feels like you're playing against a monkey and I have a strong rule on how many unga-bungas I tolerate in my maches (which is one - ME!). Dont get me wrong, I would LOVE for them to do it right and the idea is certainly a step in a right direction, but so far it does not feel like I'm fighting against an actual person. Almost as if "being a real player" goes beyond the capabilities of AI to make a statistically most appropriate choise. 🤷♂
Your conclusions are correct. It’s just sad that because there’s zero logic to it, there’s actually zero learning, so it’s just a 5 minute gimmick and not a learning tool for beginners, intermediates or pros. Even a beginner will notice it will eat spammable simple strings, and even if you ”teach it” that it should block and punish that string it still doesn’t learn to do how you reacted to that situation. I just spent 3 hours trying to teach my ghost to block Reina’s DF 1,2 but it just ate it 2/20 times, even if i blocked and punished the same move when it tried it. So you can’t teach it anything literally. Even if the AI doesn’t use logic per se, it should actually learn what the player is doing against simple move spam. It should be so that it should notice i’m getting hit by this move 18/20 times.. let’s use this move against the other person to see how the human reacts to it, and then adapt to it with the same percentage of success as i’m blocking and punishing it. This way you could teach the AI it’s weak points and it could slowly learn to be much more like you without real logic, just by using the same thing it’s doing now but would actually learn something useful. It’s way too slow to learn anything useful if it can’t learn to block a single two move string in 3 hours. So it’s just a gimmick sadly. Imagine if it learned on the fly with some sort of logic that’s just tied to your decision making and reaction times and learned your logic without becoming an impossibly good AI. It would be a great learning too even for pros because it would find weaknesses in their games too and could give them new ideas in how to deal with situations.
I found a way to break the ghost. If I pick Nina and start looping d4,1, it never manages to escape it. Furthermore, it never loops the same offense on me. I have seen other ghosts of higher ranks try and deal with it, but I never got my ghost into a situation where it could learn how to deal with it properly. Its ability to learn is a bit inhibited in some way. This is all from the demo.
Whatever, Ghosts has always been a gimmick ever since Tekken 5 PSP days. Now, if they could make AI more difficult in single player mode, that would be great. Fuck the trauma from playing online.
I'm a Tekken intermediate at best. My friend is a complete fighting game casual and was fighting my Kazuya ghost on the demo. After losing 2 rounds against it, he spammed his favourite move in Tekken............ UF 4,4,4,4 with his own Kazuya. And came back 3-2 because I rarely if ever use the string. My ghost just did nothing 🤣
I've played with mine pretty extensively in the demo, its not perfect by any stretch but it's a massive step in the right direction in fighting game ai, I can catch it out with stupid things but it also learns the counters when it tries it on me and I.punish it.
While I cannot deny that it's a fun addition to the game, the A.I just never defends well, regardless of how much it copied your offense and on top of that, it doesn't know how to handle rage arts what so ever, it will almost always attack you when you do them, usually leaving it very vulnerable to them and it never use the rage arts themselves, which crippled it severly, I was baffled by how I was able to dominate the super ghosts with my terrible Nina, considering how badly of an @$$ whooping I got during the closed beta online.
Ghosts are overrated. The best way to learn and improve is to actually use Practice Mode to its full advantage. Offense training - Go through your character movelists, find combos, etc. Defensive training - Set your opponent to do one action at a time, go through their entire movelists, learn how to recognize moves and punish (it will come naturally eventually), etc. Then try combos, movement, etc. You can even take moments from your replays where you got bodied and learn how to counter that. Lots of tools. Ghosts will never teach you why your moves work the way they do. They're just a thing to mess around with. And they're nothing new. You could send and receive them in Dark Ressurrection on the PSP too.
![Top 100 OP TEKKEN Moves of All Time [ft. Joey Fury, MYK, That Blasted Salami & More]](http://i.ytimg.com/vi/yLUnjjQ-OcE/mqdefault.jpg)
Hi guys, I saw a lot of comments being really happy that I'm "back". I really appreciate that, but I want to be clear I'm not coming back to making consistent videos. While I'm nowhere near the scale of TH-camrs like Tom Scott, I felt similarly burnt out trying to make regular videos. From now on, I'm simply going to make a video when I get an idea I'm motivated enough to work on, keeping TH-cam as a hobby instead of as a side hustle. Thank you guys for the love on this video, I'm very excited for Tekken 8 so you should see more of me soon ❤
Jesus loves you
My ghost played really stupid... mostly because I'm bad at tekken
sounds about right
@@Daemon1995_ practice makes perfect😂😂😂
My ghost is stupid because it stops adapting at a certain point, I've started rounds almost exclusively with down-back 3 as Jin and my ai hasn't done it once. Among other things
Lmao
This is definitely how my ghost will play lmao
regarding the overwriting of old data with new data:
Eddy Wang from Damagexcaling tested and confirmed that the ghost can only learn a finite amount of data, after which it will not learn anything new
he found that in order to learn new data, the old data must be deleted to make room for more. we're all hoping that the full release will resolve this issue by allowing old ghost data to be overwritten, but that does not appear to have been implemented currently
It makes sense to limit it in the demo. They could really do any number of things with the ghost feature on release and in the future. Really looking to see how this feature..... evolves.
That actually makes sense. I spent most of my time teaching my ghost giga shitpost tactics. But then I couldn't get it to learn how to run away at low health, get a large space, ki charge into rage art when they run ip to me. I guess I filled it's data already by that point.
He also pointed out how the ghosts have their own seperate rank display. And how they're all locked to 'beginner' in the demo. Maybe the limit of how much data is stored increases with rank
In Tekken 6 after every time you quit to the main menu offline it will ask you if you want to overwrite the ghost data....
So I'm sure T8 will have it.
T6 ghost was/is very good. So if it's better than T6 ghost we will be very good. T6 ghost will do sidestep counter hit setups and full combos with wall carry to wall with wall combo. So yea....,.
Thanks for the insight. I like the approach and concept of including such a ghost into such a game
During the LCQ, one of the comments in the stream said "I don't know who to root for here so I'm going for Shirdel since I like his videos". That made me find your channel, and man this video was amazing. Love the analytical style, production value, and entertainment factor. Looking forward to more deep dive content like this for the full release!
My hypothesis for the part in 9:05 (Since I studied Machine Learning and deep learning) is that there are weights (a random value) that determines the importance of the current node. The nodes with higher value are most likely kept and not discarded.
Also another observation of mine, I played a lot against my Kazuya ghost without using DF3,2,1 string.. So the ghost never used it before, but when I try to do the mind game by holding the last input of that string to get plus frames, every single time my ghost punishes it perfectly with 112 and holds his block if I don't hold (Every single time) (Even though I didn't explicitly teach him to do that by doing it myself before) So I believe the ghost learns the frames of the moves, so it will punish any -14 move with the same punish you used for any other -14 move.
To add on, I also believe the ghost learns combos near the wall (Although it took a long time before the ghost started doing wall combos correctly).
Also, The ghost doesn't learn the sequence if the sequence ends the round, It took quite some time to purposefully use rage art in situations that wouldn't end the round in order for the ghost to learn to use rage art.
That's really interesting stuff! The last part about not learning if the sequence ends the round makes a lot of sense, that seems like a huge oversight to me. Regarding df321 mind-games, that could be related to the different states. Because the charged move would count as a different state, the AI might be immediately recognising that and punish you. See what happens if you don't go for the last hit at all.
Yeah I noticed this too, the tell tale "learning" notification.never appears at the end of the round either.
I also had the experience of the bot punishing forst hit of a string (on a different character) that I never used before
There's a GDC talk "Designing AI for Killer Instinct" which describes the similar Shadow AI in that game. Their system matches your assertions perfectly - it performs no extrapolation. What they do is copy inputs from a replay that best matches the current situation. The situation is measured with hand-chosen metrics, and the best match is picked with a nearest neighbor algorithm with some randomness.
Finding analytical, very well produced videos like this for fighting games in general is extremely rare, and I'm very happy you're filling that gap. Keep up the phenomenal work!
Can't wait to see what players and the devs do with this feature in the future.
I have a feeling that, if similar, new data is preferred over old data, simply because this mimics most players getting better as they play more. It'd be interesting to do a multi-hour test where you play to your best, folllowed by playing distinctly worse and seeing if it prefers new data. Or we ask namdy bamdy.
this was a super interesting video. I hope the developers will talk about this more. I love learning about and interacting with AI.
it never ceases to amaze me how developers just refuse to improve enemy AI.
even in the one genre where the bots are literally meant to mimic human opponents and you only have 1 entity to work with, they just don't.
if you're lucky they'll read your inputs and call it a day e.e
This is definitely a huge step in the right direction but I absolutely agree and have always complained about it.
Hopefully the ai in tekken 8 isn’t horrible
If you try to learn game AI programming and it will cease to amaze you lol.
Loved you in the LCQ, amazing run. Not only are you a great content creator, you’re also a very good player.
The ghost feature has been incredibly interesting to me, i had a feeling it worked like you explained but didn’t have the knowledge to put it into words, really loved this deep dive and i hope to see the ghost feature further developed over time
Do you think the ghost data file size limit of 8321 KB is preventing it from gaining complexity? I noticed that the ghosts of top ranked players weren't very complex and were easily defeated by hyper-aggressive moves, and I saw that the downloaded ghost files are all 8321 KB. Could this be an artificial limit placed by the developers, or is it developed such that ghosts only have 8321 KB to work with? I've only tried a little bit of machine learning, but I found that neural networks for even simple games like Breakout or Pong could have larger file sizes than Tekken's, so I wonder what could happen if it had more space.
I'm not sure how accurate all of this is, because I've been fighting my own ghost lately, and I swear that the ghost uses some moves that I have never, ever used myself, or if I have, they have been accidental, so they must have some form of agency.
this is still infinitely better and much more humanlike/entertaining than input-reading & makes me wanna play tekken single player modes for sure
actual learning takes human hand crafting to build in algorithms & on how to notice data patterns & how to respond to it. That requires a different type of coding.
based on a very strong player is to have a bunch of matches. that'll be very valuable.
Really sharp and insightful analysis - good to see someone really getting into the weeds on how the ghost system works.
I'm a reinforcement learning specialist. I am quite intrigued by the Q-learning approaches of super ghost battle.
If you wan we can have a talk about possible technical details of the implementation.
This AI sounds pretty entry-level to me. Which, seeing how this mode is an extension of Arcade Quest, makes sense. An introduction to the serious play style of characters for casual players that are easier to digest, as an online player that's stomping them out can easily switch to mix ups before they can adapt. I can dig the idea, because the bridge from casual to competitive still doesn't have the clearest path.
I just wish it didn't remember sub optimal punishes so harshly, because most times I'll do a jab punish for a mix-up, or, honestly, laziness.
Is the ghost AI smart enough to develop rollback?
Top tier comment
The game has rollback
Non-injective...
Surjective...
Can't believe that I finally come across something I learnt in a class through Tekken.
My Ghost AI caually dropping combos as me im so proud of him
A point i want to share regarding question one, the ghost doesn't choose randomly but will prefer your optimal combo considering space between you and him and you and the wall, i saw that depending on how much i was away form the wall he preferred a combo that putted me on the wall instead of the most damaging without wall carry.
Another thing that blew my mind is after watching my friend and trying it my self woth kazuya, also the ai consider how much health it has left, if a player hal low health and because of that gets more nervous dropping more combos, the ghost will drop the combos as well when at low health or stressfull situation, it would also do some kind of preference in okizeme or standsr situations as well as reflexes in matchups you never did, it was pretty rough when we tested it but it tried to do mimic to an extend, we only tested it for like 30 hours not even being that good in tekken, pretty little time to see how much potential it has
the legend returns. was a huge fan of your pro tekken history history series and its nice to see the analytical videos make a return
this video could be like half the length and it'd still get all your main points across.
Welcome back - bring back pro Tekken history! ❤
Look Gary, there I am!
I main Yoshimitsu, he's been my main since 1995.
Even though I have spent a good amount of time training my ghost, and using every possible move that yoshi has, the AI can't seem to learn past a certain point. I can do a few difficult combos, repeat it often, and it just won't learn it.
But it's okay, it's still an interesting concept that I'm sure will take time to get better on the developers side.
1:05 KI '13 shadow clones were likely a precursor
Didn't even know this channel existed. Subbed, nice content! Upload when you feel like it man, no pressure.
This was exactly my feeling after spending time with my Ghost. It's just a replay machine. It doesn't actually replicate my style beyond a replay.
Eh, not sure about that, my ghost was pulling shit I never did but would 100% be something I'd do
Take a look at the video from GDC “Designing Shadow AI for killer instinct” - its absolutely brilliant and I’ve been theorycrafting on my own for ways of implementing this in a smash-like game. The Shadow AI from that game has the same idea as the ghost AI from this game
spent a few hours last night and today and it started learning my new mistakes and combos lmao so looks like it does overwrite some but also remembers some old combos
I kicked the shit out of my ghost over and over, mainly because I'm terrible on defense, but the one time my ghost had enough in the tank to throw it all back and win......it felt like I was playing against a little brother who finally picked up what I was doing rather than myself.
But that was still worth a lot to me as I immidiately hit Rematch.
I see this mode as a leaning tool for people who hate to go into replay and rewatch their games. Its a fun mirror to see what you yourself are doing. For this purpose its very good.
This is exactly the video I was hoping would exist on this topic holy shit
Bro, have to say, I'm happy to hear your voice again, it's been a bit too long 😄
Great video!
I hope they release an update in the future for all the improvements you mentioned at the end
Eyoo welcome back, great showing at the LCQ!
Man I hope the next update allows your ghost to fight another one of your ghosts, like imagine a pro ghost vs another pro ghost. I honestly dont like the cpu thingy as the actions are too slow, ghost vs ghost is quicker and much more better (hope they add it)
Awesome vid mate! Really informative!
10:22 THIS. This is the bane of my existence. I'm playing Jin not even day 2, I do Electric Thrusting Uppercut but whiff the EWHF and do df2 instead and the ghost will NOT let me live it down. It barely does the correct combo no matter how many times I do it.
Its a shame the ghost can only effectively learn mirror matchups. But I also laugh whenever my kazuya perfectly duck launch any 1,2,4,3 attempts I accidentally throw up.
the ai moves isnt permanent like the more u play with your ghost the more it addapt? or if u quit the round itll reset its progress?
This video was very intriguing to watch and it made me think of Super Smash Bros Ultimate's Amiibo training. It's rather similar to the Ghost and I think it might be interesting to look into on your own time as it varies enough just from the differences in games alone.
@shirdel the Killer Instinct reboot from 2013 (its on steam) has something called Shadow Lab that allows you to create an A.I. based on your replay data. these Shadow A.I. are just as advanced as Tekken 8's Ghosts, if not more advanced (they will literally learn to teabag), and in fact theres an hourlong lecture from GDC on youtube from a former dev of Killer Instinct about the creation of Shadow Lab mode!
holly hit this video is top tier prt 2 is in the making there.s no doubt about it
So, I pretty much only play old tekken games. Which means that my fundamentals are really solid, but a lot of the new tactics based around air juggles and wall carry and wall play I never really learned, and am getting the hang of now. It really feels as though my ghost grasped the importance of these secondary tactics right around the same time I did. It overwrote old habits very quickly, and continues to do so. I would say that this either means the routing is very good, or it can start calculating what decisions lead to the overall highest damage output.
I like how one time when I played with ghost I used a rage art in a combo awhile back. Then later when ghost launched me it was in rage and did the same combo ending in rage art.
welcome back bruva. We still need them pro tekken history videos :)
I had my AI ghost use a grab despite not having used grabs yet on it
Omg I'm glad you are back, love you man
I’m a Tekken Vet & I’m really good at game 3 mains in red & purple ranks
I love the Ai because it forces you to beat yourself thus making your main character better in execution & defensively
I wondered what they would do if i stayed on the ground and the ghost literally doesn't know what to do
To me a HUGE issue is that you have to fight your ghost for that. It would be better if it learned as we play, but so far, you need to go to a specific part of the game for it to train.
The game itself says that just playing other modes will help it learn. So it will probably be more obvious in the full game
@@Romapolitan ohhhh ok, I did not know that! Thanks!
One obvious flaw I found in the demo was it didn't properly learn to duck a high string.
At the same time I was amazed by their small SS whiff punish. I was like no way they learned that from me 😂
If machine learning is used. My guess is as you play the game it makes predictions of how you'd act while you are currently playing. If it's correct it's rewarded.
Ghost has already crossed a threshold where it plays "player-like" which is enough for me. Sometimes I can't tell if I'm playing a ghost or a player. I knew right away that it wasn't a very advanced AI, but it's worlds different from the old type of dumb-AI that their CPU used to use, where the most it would do is 1 random move at a time. We're entering a new era of AI opponents in game design and I'm honestly quite excited for it. I can't wait for someone to make the next soulslike (hate that term but for the sake of brevity) type game except the bosses fight almost like PVP players would.
My bear bit the hand that fed him.....
You gotta remember, they would want the AI to optimise its options as much as the individual does, not to play using your techniques but without your level of thoughtfulness.
It's true that they have failed to implement intent, but they would want to copy the players intent, not optimal intent.
If I press on minus, my ghost should do so too
What does Shirdel mean by the question "Does the ghost have agency?"
OK, but important question. Will it know your inputs after you win a round to sayyy ki charge, Miguel's f3+4, or Raven's labyrinth over and over again? This is vitally important.
No it does not register after after one side is dead
I really like this feature. While I understand frames and punish windows, I rarely use practice mode to learn to beat a certain characters, it gets boring. I rather play with or against the character, observe and adapt.
With this system I can learn a character and when I pick a new one I can practice by playing against my AI of the previous one to skyrocket my overall knowledge of the whole cast by repeating this process
The funny thing about the ghost system is that this isn't the first time they had it. It was in T6 back in '09.
I'm not sure precisely what's going on, but I have noticed the ghost will occasionally do things I have never done. For example, I was playing Kazuya and did not one wavedash, EWGF or backdash cancel and the ghost was occasionally pulling them out.
Welcome back dude!
Just started messing with mine.
Progress:
1st match, cake
5th match, started using my combos
10th match, learned heat & rage
20th match, used moves I didnt know😅
30th match, blocked all i threw at it, theee in some grabs
40th match, counters my grabs 🥺
57th match, at this point its a fight to see who moves 1st...Ive played as Jin for yrs so I know how he fights. My only saving grace at this point is i target the frames where he cant block😂 i fear whats to happen next. I stopped teaching it
When I attempted to do a wave dash combo the ghost picked it up the next round meanwhile I'm still getting it down 😂
is this in the demo, or do you have an early release copy? I'd love to test this shit out right now.
in demo, win the tourney at end of arcade mode to get it
no matter how many times i ki-charged after winning a round, the ghost never copied it, obviously it can’t be that smart
I'm no AI expert, but I'm a regular software engineer ...and my intuition (which might be wrong) is telling me a few things.
1. The file is probably always the same size because it's just a big list of weighted outputs for every predefined input. New data doesn't get "added" in the traditional sense -you're just changing numbers.
2. I think it's only using a certain amount of your most recent matches, as to reflect your current playstyle more accurately.
3. There's probably some basic AI agency, but only a Very Easy bot or something that things are layered on top of. I have a feeling it would need way more than a couple of matches worth of data to be able to play deterministically.
Is it any different now, that the ai can learn from all your ranked games?
with that british accent, i think you're smart enough that whatever you say is a fact
So far I'm a quite impressed on their offense but their defense leaves something to be desired. But I suppose much of that has to do with many players going full offensive and rarely blocking, especially against the CPU. There's no true replacement for playing against real human players but the ghosts are still more fun to fight against than the default AI.
I have hypothesis about how ghost ai dumps data and this will also solve why game data is not expanding expotentially.
"If player forexample makes the best combo punisher 0% and 25% of time he makes some combo punish."
Here I predict ai would make 0% of time the best combo punisher .
Then he plays more and gives more info to a.i.
In the next example.
"If player forexample makes best combo punisher 25% and 50% of the time make some combo punish."
Here I predict ai would make best combo punisher 100% of the time or maybe there is some variance adjusted?
In time this results and correct button presses are like nodes that get activated and the most sophisticated ai ghost will never make input mistake.
In this model of ai. The final best ghost could same amount of info or close to it whether the switches are on or off. So you save hardrive space. I am currently only advancing to test this ai but I am in terested in machine learning.
Thank you for your time! Have a nice day.
could it be that there's a finite number of fighting options and your "ghost" is scored on all of those options. (Given a discrete scenario you throw sometimes. So, you establish a 4 out of 10 probability of throwing in that scenario. That "throw score" of 4/10 would mean that sometimes you throw... so when the ghost is in that situation, it throws 4/10 of the times ) The hard part about building ghost intelligence like that is identifying the scenarios and scoring the action. Refinements in that system could be thought of as "advancements in the intelligence".
So, basically the ghost is just a set of values for pre-determined (yet undisclosed) parameters? Every check it does it asks itself a question "what would be the best probable move to do after, given the current state", which it gets directly from the current state of the game mode (ue thing). I think that THATS THE EXACT RESON why it feel like your ghost does some "CPU"- level stupid stuff sometimes, especially when you do nothing, leaving it suddenly with little gamestate changes. My guess is that in the absence or in cases of low projected probability, it just resorts to its next move being "the default" for conventional decision-tree based bot. Sprinkle some salt (random-ish variation) and boom!
I also think you're giving bamco more credit than they deserve when it comes to this thing. I highly doubt they've engineered a custom-tailored neural network just for the purpose of FG ai. They've probably mapped out all the moves (because they have to do it anyways lol) and states of the game (time, position, juggle, heat amt etc.), and used some pre-existing network to "teach" it with the game data as well as made it being able to execute stuff as a cpu-controlled pawn.
In the demo, it feels like you're playing against a monkey and I have a strong rule on how many unga-bungas I tolerate in my maches (which is one - ME!). Dont get me wrong, I would LOVE for them to do it right and the idea is certainly a step in a right direction, but so far it does not feel like I'm fighting against an actual person. Almost as if "being a real player" goes beyond the capabilities of AI to make a statistically most appropriate choise. 🤷♂
IS this the same like T5 T6 TTT2? Like Arcade ghost battle?
Your conclusions are correct.
It’s just sad that because there’s zero logic to it, there’s actually zero learning, so it’s just a 5 minute gimmick and not a learning tool for beginners, intermediates or pros.
Even a beginner will notice it will eat spammable simple strings, and even if you ”teach it” that it should block and punish that string it still doesn’t learn to do how you reacted to that situation.
I just spent 3 hours trying to teach my ghost to block Reina’s DF 1,2 but it just ate it 2/20 times, even if i blocked and punished the same move when it tried it. So you can’t teach it anything literally.
Even if the AI doesn’t use logic per se, it should actually learn what the player is doing against simple move spam.
It should be so that it should notice i’m getting hit by this move 18/20 times.. let’s use this move against the other person to see how the human reacts to it, and then adapt to it with the same percentage of success as i’m blocking and punishing it.
This way you could teach the AI it’s weak points and it could slowly learn to be much more like you without real logic, just by using the same thing it’s doing now but would actually learn something useful.
It’s way too slow to learn anything useful if it can’t learn to block a single two move string in 3 hours.
So it’s just a gimmick sadly.
Imagine if it learned on the fly with some sort of logic that’s just tied to your decision making and reaction times and learned your logic without becoming an impossibly good AI. It would be a great learning too even for pros because it would find weaknesses in their games too and could give them new ideas in how to deal with situations.
Speaking of ghosts, hey shirdel.
I'd like to see a comparison with virtual fighter 4 evolution's ghost system
The more technical you play,the easier it’ll be to copy you. The more emotion and instinct you play with the less like you it’ll be.
Me encanto mucho esta genial❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤
I found a way to break the ghost. If I pick Nina and start looping d4,1, it never manages to escape it. Furthermore, it never loops the same offense on me. I have seen other ghosts of higher ranks try and deal with it, but I never got my ghost into a situation where it could learn how to deal with it properly. Its ability to learn is a bit inhibited in some way.
This is all from the demo.
not sure how smart the ghost in demo were but some of them couldn't not low parry everything and basically be psychic.
Whatever, Ghosts has always been a gimmick ever since Tekken 5 PSP days. Now, if they could make AI more difficult in single player mode, that would be great. Fuck the trauma from playing online.
Tekken 5 A.i be like: first time??
I'm a Tekken intermediate at best. My friend is a complete fighting game casual and was fighting my Kazuya ghost on the demo. After losing 2 rounds against it, he spammed his favourite move in Tekken............ UF 4,4,4,4 with his own Kazuya. And came back 3-2 because I rarely if ever use the string. My ghost just did nothing 🤣
Have been playing Tekken 6 and the AI was something else, they should just replicate that.
It’s like a digital TEKKEN en amibo
Can someone tell me how to add more ghost? I only have one and that's Leroy, i can't find where to add more ghost 😢
Fight your Leroy ghost with a different character.
Just a question to the people that played it for some time.
Is the Ghost/AI impressive?
I've played with mine pretty extensively in the demo, its not perfect by any stretch but it's a massive step in the right direction in fighting game ai, I can catch it out with stupid things but it also learns the counters when it tries it on me and I.punish it.
New Joel like devil Jin
I wish i had this in T7 to fight my Negan for shits and giggles.
While I cannot deny that it's a fun addition to the game, the A.I just never defends well, regardless of how much it copied your offense and on top of that, it doesn't know how to handle rage arts what so ever, it will almost always attack you when you do them, usually leaving it very vulnerable to them and it never use the rage arts themselves, which crippled it severly, I was baffled by how I was able to dominate the super ghosts with my terrible Nina, considering how badly of an @$$ whooping I got during the closed beta online.
Ghosts are overrated. The best way to learn and improve is to actually use Practice Mode to its full advantage.
Offense training - Go through your character movelists, find combos, etc.
Defensive training - Set your opponent to do one action at a time, go through their entire movelists, learn how to recognize moves and punish (it will come naturally eventually), etc.
Then try combos, movement, etc. You can even take moments from your replays where you got bodied and learn how to counter that. Lots of tools.
Ghosts will never teach you why your moves work the way they do. They're just a thing to mess around with. And they're nothing new. You could send and receive them in Dark Ressurrection on the PSP too.
Not available on PS4.
3:23 bro could have just used a star as a symbol
I wonder how it compares to KI or Dimensions' ghost AIs.