There was an interesting talk last year at GDC about the machine learning Riot was trying for TFT. One of the biggest issues they ran into is getting TFT to run in a headless state (without all the graphics and such) so that the ML could run a game of TFT in a short enough time frame to actually get enough data. It would have require so many changes to how the LoL and TFT engine was coded, that they ended up creating a python sim of tft instead. Unfortunately that talk is still behind a paywall until 2025, but you can find the slides on the gdc vault if you're interested in the approach they took.
there definitely are an insane number of variables to create an actually perfect ai, ranging from positioning, leveling, gold/econ, rolling odds, augment considerations, lobby comp playing, and having them all be interconnected only increases the compleixty. Many of this complexity is simplified in current meta to a rough this is how to do it, but its only because game to game it is too difficult for the average player to reasonably account for all the factors and instead can only rely on "luck"
People have to realize, TFT has an absolutely god damn insane amount of variables and permutions, so much so that using Ai to balance isn’t just as easy as “duh plug the numbers in and it will tell you what is OP” and anyone who thinks like that doesn’t understand machine learning.
I think anyone who thinks like that doesn’t understand AI, but they don’t misunderstand ML. Obviously it’s not so easy a caveman can do it but zero effort to even try outside a lone attempt back in set 3 is wild
@@spencerweiss6337 I had a more detailed response in another comment, but I really don't think it's wild. I think they tried enough to realize that this would be a massively expensive and difficult task with very, very low return on investment. Balancing TFT "good" doesn't exactly rake in the big bucks, and even there, good is in quotes because good balance is most relevant in player perception rather than computer optimization.
@@BlanekdCheque fair enough; it’s a risk for sure. I’m far from read up on much of the recent research that’s been published over the past few years but there’s been a significant amount of time lapsed since their last attempt and I gotta imagine they’d be able to develop something excellent that would prove financially worthwhile. Especially if they’re feeling challenged by the complexity ceiling rn
@@BlanekdChequeExactly. Even if you spent the resource and finally got a machine that was significantly better than human players, balancing for that hyper optimized level of play would be wholly irrelevant to you actual playerbase, because no human would actually be playing on that level, and even if they were, metas often have more to do with trends than actual strength.
@@spencerweiss6337 Ehhhhhh, both developers and server time is at a premium right now. Maybe I’m reading too much into it, but I really don’t think the issue is “Too complex” but “Too complex to be worth it”. Also, AI projects like these (OpenAI’s Dota bot, AlphaGo, Stock fish, even ChatGPT) are largely unprofitable ventures. They have potential, and they have advance the research, but making AI suddenly profitable is not a trivial task. Especially if at the end, all you have is a good balancing bot when your company makes money from skin sales.
This patch has to be the most balanced one we had in years. I mean, you could play a lot of good comps. It's just a matter of how contested and lucky you are.
Ive played alot of auto battle games. And imo TFT is one of the hardest and complicated ive ever played. Ive played auto chess, magic chess, chess rush, dota underlords, and the one with knight theme with alot of people on low health. And yet i cant even push to plat this set, ive reached mythic on magic chess which i tought magic chess is harder, since the tactician has skills. But no, as my experience TFT is far harder than auto battler ive ever played, theres just alot of variable to play, like augments, synergies, artifact and in set 11 theres encounter. I cant imagine how hard they can manage to make it balance, or just enough.
People who are saying it’s easy and that Mort underestimates AI. You have to consider that if you make a lobby of 8 AI to play each other millions of times, this will not translate to 8 humans playing like humans if both experiments were run blindly and so balancing around what the AI does wont matter unless humans get access to the AI games and learn how they played a specific set / patch.
well chess is complicated lol. chess has lots of moves, but its siple, you just need to corner the king, tft is complicated, you need to know all builds, counters to them, where to put units and much more, thats dificult tu configure, may as well bild an AI from scrach and sell that lol
despite chess having ai the game is really only "solved" if there's like 8 or less pieces on the board, all the engines are capable of missing "best" or better moves. tft is way more complicated than chess so i would say a tft ai is a farway off
Machine learning developer here. In terms of permutations chess probably has more possibilities. What makes tft infinitely harder to solve is the lack of turns. It wouldn’t be hard to make an ai roll 100 gold instantly but positioning is a constantly changing variable. If you know who you are fighting it’s a lot easier but normally it’s between 3 boards and an AI would have to weigh how each adjustment changes 3 potential fights. Furthermore, in a perfect world an AI would probably keep different board permuations on bench(sage lillia vs 7 mythic lillia). Each board permutation would have pros/cons against 3 potential opponents which up until the last second can change. TLDR: the lack of turns in tft make it extremely difficult for an AI to find a linear solution(guaranteed checkmate in x turns) and the order in which actions should be done in can’t really be determined
TFT has so many more permutations, its not even funny. In terms of raw permutation numbers, the difference is so insanely big, were probably incapable of even writing it down.
@@789blablajaja Chess engines run on tens of years of cummulated experience already, on top of chess being the same game for hundreds of years. If you would make a bot play on the same patch, it would probably be able to easily climb to high elo after a year or so of training, maybe even less. If there were AI that could curb stomp pro teams in DotA years ago, I'm sure they can manage to consistently win in TFT, the effort is just not worth it for the sake of balancing.
Machine learning is not only reinforcement learning. MetaTFT has machine learning and not reinforcement learning. If you played Boardle and considered in anyway, the right most statistic - win rate for bot or top board - you have then utilized ML to learn about board strength (and might affect your view on the current META). With the large amount of data riot has, surely there is some system that utilizes ML to help discovering comps, builds, or even anomaly detection (great correlation to bugs).
Though to be honest, I’ve read more and more about Riot’s company value and software dev roles. Riot would probably be open for third party tools, rather than building internal ML pipelines that are prone to simulation errors like he mentioned with Iniko
Well tft is not balanced for sure. If you just follow top tier you would get in 4th place. But you would lose the game otherwise even if you complete the comb
I think the issue Mort pointed out about not having the stats before release could be resolved with A.I. Rather than making it build the sets you could have it learn the game and the set you want to play test and have it run thousands of games each day (I imagine it would be very difficult to get it to play in an "intended way" with so many variables each game, but achieving that would be huge), adapting with the knowledge of what works and what doesn't and then having actual numbers to work with. I'm not an expert on the matter but this seems like a logical approach to me, would love to know if I'm missing something.
Its not that TFT is more complicated. Its there is too much entropy for AI to strategize Also... TWITCH CONFIRMED FOR SET 12? ua-cam.com/video/2An0vMRJKMU/v-deo.html
Your game is evidently too complicated for all of you nerds as well… as the balance sucks to this day, and it trickles down to 2-3 comps at the end of each set in the last 7 sets.
wtf are you even yapping about? xdd You, with zero software/coding knowledge claim to know more than the entire dev team that have been working relentlessly on this game for 5 years. it doesn't even matter how good you are at the game, like I don't even care if you're dishsoap, that's just a brainlet take.
"complicated" is measured differently for AI or software in general. It has more to do with what types of models you can apply and how much information (and how many different kinds of information) it needs to process... I'd say you can make a TFT bot about as good as OpenAI's DotA bot, but to get a bot good enough that it's genuinely useful for development? I'm willing to guess TFT is wildly more complex
it's also how much you want to put in resource towards it instead of other things. I'd say that it really wouldn't be effective unless you put a lot of resources to it, which kind of defeats the purpose of building it.
Dota is much easier to make AI for because a lot of the complexity comes from awareness and execution. Both things a computer has easy access to and can isolate. One of my favorite moments from this set was watching bard /tahm get nerfed because everybody in the lower elos didn’t find solutions for it. Only for the nerf to get reverted and instead buffed 3 patches later. A perfect AI (which could take YEARS of testing and changes very set) would have found those solutions to bard very fast and BUFFED him because he couldn’t deal with shen frontlines, heavenly yone, ghostly damage ramp and fast 9 legendary boards.
i'm sorry but set 3 was 4 years ago and I'm pretty sure ai has gotten way smarter since then, also, you had >4 years to develop a better one. love mort, but he sounded proud to not be using the most powerful tool us humans have right now to enhance his game
@@xMetixYT if white was OP, then 50% of grandmaster games wouldn’t be draws, and 75% over 2750 wouldn’t be draws. White has an advantage, not OP. I could beat 95% of players with black blindfolded.
@@dementedmanatee1298 Added complication does not imply skill differentials if unbalanced. And if a 2024 coding team for a trillion dollar company can't get a machine learning model to understand their game, then it is either unbalanced, non user-friendly, or both. Btw, I worked in UI and animations for game companies for 7 years. Primarily chess based.
Wow... Just wow... I've seen mort say some really out of touch things but this is just out of this universe levels of delusion. The amount of datapoints to simulate TFT is comically low in comparison to what many commercial applications are doing today. Yet another example of Mort's ego standing in the way of the game becoming better. Now I understand why this game is constantly in a state where half the units are un-carriable trait bots (kaisa right now) and there's really only a pool of 10-20 viable comps.
If you try to create a model for the game, you'll soon realize it's very difficult to create an AI model for the game that's not overfitted. Also, just completing the machine learning training takes months since you have so many iterations and edge cases to run even WITHOUT accounting for the numerous balance changes every two weeks😂. Using machine learning for a game like TFT is such a niche topic and definitely not worth the time and resources you would have to invest to actually make a successful model that'll probably change in 4 months' time when they release a new set 🤭 Sounds like what you're experiencing is a "skill issue" and you should just "play better"
There was an interesting talk last year at GDC about the machine learning Riot was trying for TFT. One of the biggest issues they ran into is getting TFT to run in a headless state (without all the graphics and such) so that the ML could run a game of TFT in a short enough time frame to actually get enough data. It would have require so many changes to how the LoL and TFT engine was coded, that they ended up creating a python sim of tft instead.
Unfortunately that talk is still behind a paywall until 2025, but you can find the slides on the gdc vault if you're interested in the approach they took.
yeah which is exactly my point of it being a resource issue
there definitely are an insane number of variables to create an actually perfect ai, ranging from positioning, leveling, gold/econ, rolling odds, augment considerations, lobby comp playing, and having them all be interconnected only increases the compleixty. Many of this complexity is simplified in current meta to a rough this is how to do it, but its only because game to game it is too difficult for the average player to reasonably account for all the factors and instead can only rely on "luck"
People have to realize, TFT has an absolutely god damn insane amount of variables and permutions, so much so that using Ai to balance isn’t just as easy as “duh plug the numbers in and it will tell you what is OP” and anyone who thinks like that doesn’t understand machine learning.
I think anyone who thinks like that doesn’t understand AI, but they don’t misunderstand ML. Obviously it’s not so easy a caveman can do it but zero effort to even try outside a lone attempt back in set 3 is wild
@@spencerweiss6337 I had a more detailed response in another comment, but I really don't think it's wild. I think they tried enough to realize that this would be a massively expensive and difficult task with very, very low return on investment. Balancing TFT "good" doesn't exactly rake in the big bucks, and even there, good is in quotes because good balance is most relevant in player perception rather than computer optimization.
@@BlanekdCheque fair enough; it’s a risk for sure. I’m far from read up on much of the recent research that’s been published over the past few years but there’s been a significant amount of time lapsed since their last attempt and I gotta imagine they’d be able to develop something excellent that would prove financially worthwhile. Especially if they’re feeling challenged by the complexity ceiling rn
@@BlanekdChequeExactly. Even if you spent the resource and finally got a machine that was significantly better than human players, balancing for that hyper optimized level of play would be wholly irrelevant to you actual playerbase, because no human would actually be playing on that level, and even if they were, metas often have more to do with trends than actual strength.
@@spencerweiss6337 Ehhhhhh, both developers and server time is at a premium right now. Maybe I’m reading too much into it, but I really don’t think the issue is “Too complex” but “Too complex to be worth it”. Also, AI projects like these (OpenAI’s Dota bot, AlphaGo, Stock fish, even ChatGPT) are largely unprofitable ventures. They have potential, and they have advance the research, but making AI suddenly profitable is not a trivial task. Especially if at the end, all you have is a good balancing bot when your company makes money from skin sales.
This patch has to be the most balanced one we had in years. I mean, you could play a lot of good comps. It's just a matter of how contested and lucky you are.
Game is so flexible is fun
Backsizes really hurt tho
Wait, what elo are you playing at?
@@MegaRob421 I was Emerald 4 last season on double up carrying my friend. Kinda scuffed 😂but still
@@MegaRob421masters 181 here and i agree with this comment.
Ive played alot of auto battle games. And imo TFT is one of the hardest and complicated ive ever played. Ive played auto chess, magic chess, chess rush, dota underlords, and the one with knight theme with alot of people on low health.
And yet i cant even push to plat this set, ive reached mythic on magic chess which i tought magic chess is harder, since the tactician has skills. But no, as my experience TFT is far harder than auto battler ive ever played, theres just alot of variable to play, like augments, synergies, artifact and in set 11 theres encounter.
I cant imagine how hard they can manage to make it balance, or just enough.
People who are saying it’s easy and that Mort underestimates AI. You have to consider that if you make a lobby of 8 AI to play each other millions of times, this will not translate to 8 humans playing like humans if both experiments were run blindly and so balancing around what the AI does wont matter unless humans get access to the AI games and learn how they played a specific set / patch.
well chess is complicated lol. chess has lots of moves, but its siple, you just need to corner the king, tft is complicated, you need to know all builds, counters to them, where to put units and much more, thats dificult tu configure, may as well bild an AI from scrach and sell that lol
and thats for one set, what about the other sets? will you reprogram the AI every set? lol so stupid...
Best game ever. Me and my friends are playing everyday.
despite chess having ai the game is really only "solved" if there's like 8 or less pieces on the board, all the engines are capable of missing "best" or better moves. tft is way more complicated than chess so i would say a tft ai is a farway off
Machine learning developer here. In terms of permutations chess probably has more possibilities. What makes tft infinitely harder to solve is the lack of turns. It wouldn’t be hard to make an ai roll 100 gold instantly but positioning is a constantly changing variable. If you know who you are fighting it’s a lot easier but normally it’s between 3 boards and an AI would have to weigh how each adjustment changes 3 potential fights. Furthermore, in a perfect world an AI would probably keep different board permuations on bench(sage lillia vs 7 mythic lillia). Each board permutation would have pros/cons against 3 potential opponents which up until the last second can change. TLDR: the lack of turns in tft make it extremely difficult for an AI to find a linear solution(guaranteed checkmate in x turns) and the order in which actions should be done in can’t really be determined
TFT has so many more permutations, its not even funny. In terms of raw permutation numbers, the difference is so insanely big, were probably incapable of even writing it down.
@@789blablajaja i mean it depends on what you call a permutation. No chess AI is brute-forcing every combination and neither would a tft
@@789blablajaja Chess engines run on tens of years of cummulated experience already, on top of chess being the same game for hundreds of years. If you would make a bot play on the same patch, it would probably be able to easily climb to high elo after a year or so of training, maybe even less. If there were AI that could curb stomp pro teams in DotA years ago, I'm sure they can manage to consistently win in TFT, the effort is just not worth it for the sake of balancing.
@@TheCagedK A permutation is a permutation... theres no interpretation. Different game State -> another permutation.
Machine learning is not only reinforcement learning.
MetaTFT has machine learning and not reinforcement learning. If you played Boardle and considered in anyway, the right most statistic - win rate for bot or top board - you have then utilized ML to learn about board strength (and might affect your view on the current META).
With the large amount of data riot has, surely there is some system that utilizes ML to help discovering comps, builds, or even anomaly detection (great correlation to bugs).
Though to be honest, I’ve read more and more about Riot’s company value and software dev roles.
Riot would probably be open for third party tools, rather than building internal ML pipelines that are prone to simulation errors like he mentioned with Iniko
Well tft is not balanced for sure. If you just follow top tier you would get in 4th place. But you would lose the game otherwise even if you complete the comb
It is a trick question, since it is a luck-based game, it is not balanced, it is dogshit gambling. There
I think the issue Mort pointed out about not having the stats before release could be resolved with A.I.
Rather than making it build the sets you could have it learn the game and the set you want to play test and have it run thousands of games each day (I imagine it would be very difficult to get it to play in an "intended way" with so many variables each game, but achieving that would be huge), adapting with the knowledge of what works and what doesn't and then having actual numbers to work with.
I'm not an expert on the matter but this seems like a logical approach to me, would love to know if I'm missing something.
I mean you literally said in your comment why it doesn't work.
Its not that TFT is more complicated. Its there is too much entropy for AI to strategize
Also... TWITCH CONFIRMED FOR SET 12?
ua-cam.com/video/2An0vMRJKMU/v-deo.html
Tft is WAY MORE complicated than chess
Your game is evidently too complicated for all of you nerds as well… as the balance sucks to this day, and it trickles down to 2-3 comps at the end of each set in the last 7 sets.
is he sure there isnt machine learning? when you first start tft your first few games have the same players in it all the time, those are bots no?
It is not the same to make an AI capable of playing tft, or doing an AI that can be used to balance the game properly because of the data it provides
tft is not more complicated than dota and openai trained on that... has to be a resource issue
wtf are you even yapping about? xdd You, with zero software/coding knowledge claim to know more than the entire dev team that have been working relentlessly on this game for 5 years. it doesn't even matter how good you are at the game, like I don't even care if you're dishsoap, that's just a brainlet take.
"complicated" is measured differently for AI or software in general. It has more to do with what types of models you can apply and how much information (and how many different kinds of information) it needs to process... I'd say you can make a TFT bot about as good as OpenAI's DotA bot, but to get a bot good enough that it's genuinely useful for development? I'm willing to guess TFT is wildly more complex
it's also how much you want to put in resource towards it instead of other things. I'd say that it really wouldn't be effective unless you put a lot of resources to it, which kind of defeats the purpose of building it.
@@nickitnite9682there's really no need to be mean
Dota is much easier to make AI for because a lot of the complexity comes from awareness and execution. Both things a computer has easy access to and can isolate.
One of my favorite moments from this set was watching bard /tahm get nerfed because everybody in the lower elos didn’t find solutions for it. Only for the nerf to get reverted and instead buffed 3 patches later.
A perfect AI (which could take YEARS of testing and changes very set) would have found those solutions to bard very fast and BUFFED him because he couldn’t deal with shen frontlines, heavenly yone, ghostly damage ramp and fast 9 legendary boards.
i'm sorry but set 3 was 4 years ago and I'm pretty sure ai has gotten way smarter since then, also, you had >4 years to develop a better one. love mort, but he sounded proud to not be using the most powerful tool us humans have right now to enhance his game
1 view in 1 minute...
shutddup
@frejlord6794 fell off
“We’re way more complicated than chess”
But, not as balanced.
Nah white op
@@xMetixYT if white was OP, then 50% of grandmaster games wouldn’t be draws, and 75% over 2750 wouldn’t be draws. White has an advantage, not OP. I could beat 95% of players with black blindfolded.
yes, that's exactly what happens when you're way more complicated lol. You know what's perfectly balanced? Rock paper scissors
@@dementedmanatee1298 Added complication does not imply skill differentials if unbalanced. And if a 2024 coding team for a trillion dollar company can't get a machine learning model to understand their game, then it is either unbalanced, non user-friendly, or both. Btw, I worked in UI and animations for game companies for 7 years. Primarily chess based.
Wow... Just wow... I've seen mort say some really out of touch things but this is just out of this universe levels of delusion. The amount of datapoints to simulate TFT is comically low in comparison to what many commercial applications are doing today. Yet another example of Mort's ego standing in the way of the game becoming better. Now I understand why this game is constantly in a state where half the units are un-carriable trait bots (kaisa right now) and there's really only a pool of 10-20 viable comps.
idk about that Kaisa part...
@@ekiPog bro obv does not play the game lmao
That guy has no clue how machine learning works. 🤣
Kaisa is good right now and even when she was bad she was never a "traitbot" lol. I don't think you know what that means.
If you try to create a model for the game, you'll soon realize it's very difficult to create an AI model for the game that's not overfitted. Also, just completing the machine learning training takes months since you have so many iterations and edge cases to run even WITHOUT accounting for the numerous balance changes every two weeks😂. Using machine learning for a game like TFT is such a niche topic and definitely not worth the time and resources you would have to invest to actually make a successful model that'll probably change in 4 months' time when they release a new set 🤭 Sounds like what you're experiencing is a "skill issue" and you should just "play better"
How can any sane human being can say tft is balanced? 3 star kobuko wins again 3 star kayn and lee sin 😂
skill issue