FuRy Smash
Smash Cadet
i believe the cpus somehow learn. several times i have seen the cpus edgehog my aether. i mean the instant edge hog where they drop down and grab the stage just before i start my aether. then i fall to my death 
-
Welcome to Smashboards, the world's largest Super Smash Brothers community! Over 250,000 Smash Bros. fans from around the world have come to discuss these great games in over 19 million posts!
You are currently viewing our boards as a visitor. Click here to sign up right now and start on your path in the Smash community!
CPU's Learn?
- Thread starter hectichobo
- Start date
hectichobo
Smash Ace
Wow, this thread is still.... here? Wow... I was not aware of that.
Sonic dash-danced against me once o.O
JCMR-Brawler64
Smash Apprentice
I hate fight with a Lv9 com with items (like Dragoon).When the com hit me and I have a part of Dragoon, my part falls.
But when I hit the com and have a part... nothing happens.
The only way to stole the part of Dragoon is KILL the com.
Is unfair
CorruptFate
The Corrupted
So im not sure about if the debate about CPU's learning is still going on or not so this might just be a waist of time, but they do and here is why. CPU's on your guy's wii's dont teabag I asume, we'll mine didn't either until my buddy who comes over started teabagging after he gets a kill, now my comps do the same. Test this for your self if you want I will try to get a vid up once my recording works (**** vista recording). Also alot of people say "These things were programed into the game" there is no way nintendo would put/let someone put teabagging in one of there games.
GrayPlague
Smash Apprentice
- Joined
- Sep 23, 2008
- Messages
- 105
Full proof that CPU's learn: I've had a Mr. Game & Watch pseudo-chaingrab me before. I abuse that all the time just because I find it rewarding, and then an opposing G&W does it to me. -_-
Jonas
Smash Champion
It's not really a "proof". CPU's do random stuff all the time, and it's not uncommon for them to attempt the same move several times, in this case, a grab.
Ceph
Smash Apprentice
LOL I noticed the crouch taunting too! That almost gives them a personality of their own O.o scary
Denzi
Smash Master
This is a crazy and unexplained force of nature!
But we'll get it eventually.
Kind of like when we figured out that volcanos only erupt when we don't sacrifice enough people to them.
lol
But we'll get it eventually.
Kind of like when we figured out that volcanos only erupt when we don't sacrifice enough people to them.
lol
Junahu
Smash Ace
My understanding of the CPUs is that, if they learn at all, they do not take on additional knowledge but rather adapt it while they watch you play.
I would assume that a CPU that could adapt would have a series of weighted tables that listed the kinds of actions that should be taken in a given situation (e.g. "Crouch" "Walk" "In Air" "Landing" "Opponent off edge" "Opponent in air" "In range of Enemy" "In path of strong projectile "Just Got a KO". Basically, all the readings that the CPU has to do to function). Whenever the player takes an action under one (or multiple in many cases) of these situations, the game would update the tables.
It would not update the tables with the specific move, but rather the various properties of the move and a weighting value to tell the CPU how often to use the move. "D-air" means nothing to the computer, but 'knockback = 80' 'base damage = 14' would because they're nice arbitrary numbers the CPU can apply fuzzy logic to. This means that the CPU could be in a similar situation, apply the variables of that situation to the numbers in the table, and come up with a different result that works better with the circumstances.
Basically the idea is;
1) observe what the player does in different situations. Reduce the actions to variables and place them in a table based on how often they were used
2) whenever a similar situation arises for the CPU, look at the relevent table and choose an action randomly. Moves higher in the table are chosen more frequently.
3) Process the move chosen with the specific variables of the situation (ie. exact positioning, momentum etc)
4) Cross reference the result with the CPU's moveset to find the action that most closely resembles the result
5) Execute the action
5a)If in doubt, FALCON PAUNCH!
It sounds complicated, but CPU players do most of this processing already so adding a weighting factor should be comparitively simple.
Though if it's anything more complicated than Weighted decision tables and fuzzy logic then I'm afraid it's probably above me. Never did get my head around neural networks.
I would assume that a CPU that could adapt would have a series of weighted tables that listed the kinds of actions that should be taken in a given situation (e.g. "Crouch" "Walk" "In Air" "Landing" "Opponent off edge" "Opponent in air" "In range of Enemy" "In path of strong projectile "Just Got a KO". Basically, all the readings that the CPU has to do to function). Whenever the player takes an action under one (or multiple in many cases) of these situations, the game would update the tables.
It would not update the tables with the specific move, but rather the various properties of the move and a weighting value to tell the CPU how often to use the move. "D-air" means nothing to the computer, but 'knockback = 80' 'base damage = 14' would because they're nice arbitrary numbers the CPU can apply fuzzy logic to. This means that the CPU could be in a similar situation, apply the variables of that situation to the numbers in the table, and come up with a different result that works better with the circumstances.
Basically the idea is;
1) observe what the player does in different situations. Reduce the actions to variables and place them in a table based on how often they were used
2) whenever a similar situation arises for the CPU, look at the relevent table and choose an action randomly. Moves higher in the table are chosen more frequently.
3) Process the move chosen with the specific variables of the situation (ie. exact positioning, momentum etc)
4) Cross reference the result with the CPU's moveset to find the action that most closely resembles the result
5) Execute the action
5a)If in doubt, FALCON PAUNCH!
It sounds complicated, but CPU players do most of this processing already so adding a weighting factor should be comparitively simple.
Though if it's anything more complicated than Weighted decision tables and fuzzy logic then I'm afraid it's probably above me. Never did get my head around neural networks.
Levitas
the moon
On the other hand, you're wrong. People have been using a TON of ******** arguments for years now that smash cpus learn. THEY DON'T, even if there is a way to make a "learning" algorithm.
CPUs DON'T, never have, and probably never will learn in brawl.
CPUs DON'T, never have, and probably never will learn in brawl.
game set
Smash Apprentice
not in brawl, but i can make a program that learns how to play a game easily. although unless i did alot of work, it would take a ridiculous amount of time to get good. nowere near impossible though, lookup neural nets for a start.
Junahu
Smash Ace
My first sentence qualified that the post was a "what if..."
Besides, I agree that CPU's can't learn.
All the procedures and data structures CPU's ever need to act are there from the start.
I simply argued that's it's possible that they change the weighting factors of the existing AI, based on the actions of the player. That isn't learning at all. It's equivilent to "He did a Falcon kick. I should try more Falcon Kicks. Oh, I can't because I'm Ganondorf. But I have something similar, so I'll try using that more often."
To argue that CPU's can't vary their actions based on the situation and the player is an insult to the very term Artificial Intelligence. Even though they can't learn, they at least have to be able change.
Edit: Actually, come to think of it. What I described was a neural network after all. Not true learning, but rule adapting. Go fig.
Paladinight
Smash Apprentice
I'm not sure but i think they are learning because now my wolfs are using wall of wolf now (which is using a barrage of bairs to force your opponent to rethink their approach), semi-scarring and its uses the psuedo-shinespike with wolf, this is too creepy i don't fight my wolfs anymore
Hyro
Smash Lord
- Joined
- Jul 31, 2008
- Messages
- 1,386
I got shinespiked by a cpu wolf. It was even on a ledgedrop. I ledgedrop so much, the cpus must have caught on, because he dropped, shined me, recovered and I died.
Junahu is partially right, CPU DO "learn", or at least give that feeling.
An example of how do the CPU learns could be of the Thinking pattern they use; Lvl 9 Gannons and Captain falcons, have the standard reaction of usually using up-b whenever an enemy is nearby in the air, so it would be correct to say that the CPU saves a new pattern for the characters, with data of players, so whenever an enemy is in certain position depending on the character the CPU is Controling, certain action is called. It´s called conditional Branch (RPG maker series anyone?) Really simple actually.
An example of how do the CPU learns could be of the Thinking pattern they use; Lvl 9 Gannons and Captain falcons, have the standard reaction of usually using up-b whenever an enemy is nearby in the air, so it would be correct to say that the CPU saves a new pattern for the characters, with data of players, so whenever an enemy is in certain position depending on the character the CPU is Controling, certain action is called. It´s called conditional Branch (RPG maker series anyone?) Really simple actually.
Junahu
Smash Ace
Didn't really understand what you were getting at here, sorry.
RPG maker really calls them "Conditional branches"? Yikes, horrible name.
"Conditional branch" is a redundancy. If there's a branch, then it's already assumed to be conditional, since that's really the point of branches to begin with.
goku21
Smash Rookie
yo i just wanted to add in
http://www.ailive.net/liveAI.html
this program is developed as a gift for nintendo btw...
perhaps its used in brawl?
someone could sent them an email and ask?
http://www.ailive.net/liveAI.html
this program is developed as a gift for nintendo btw...
perhaps its used in brawl?
someone could sent them an email and ask?
Mario_ 101
Smash Ace
I played with a level 9 cpu today to help get used to no lag (i've been on wi-fi a lot lately) and i noticed it did SH double aerials with mario in a dittos match. O_o
Y.b.M.
Smash Master
There programmed to react to you a different depending on where you're located.....in my opinion...
estion11
Smash Apprentice
NOW
Smash Cadet
i think if a cpu gets owned depending on its position when it was killed it probably wont be there again
GofG
Smash Champion
Nintendo did not give AIs any way to learn. Sorry, it's just not true. They have a certain set of reactions to certain circumstances, and this never changes.
paulleh
Smash Apprentice
omg. i thought CPU's learning was my imagination.
every time i face a falco he always wall jumps off the ledge then side-b's back on to the stage like i always do.
every time i face a falco he always wall jumps off the ledge then side-b's back on to the stage like i always do.
ShadowLink84
Smash Hero
Bad test is bad.
Doing one test does not prove anything definite.
There are also somethings that the AI just won't do as well.
Anyways I noticed after playing several ness' that the AI does Bair a bit more often.
Nothing conclusive as of yet though.
okiyama
Smash Ace
There was an iteration of either soul calibur or street fighter for the arcade that did have CPU's that could learn from players. They boasted about it though and I think that if nintendo went to the trouble of making them learn they'd talk about it.
Sudsy86_
Smash Ace
They can't learn, eh?
How about when I word it "they can adapt"?
I play against cpu players a LOT and one thing happens consistently: when you start to take control of the match ( i.e during a long, dragged-out combo), the cpu player will do everything it can to run away and camp--including MK. That indicates something exists making it aware it is in danger if it doesn't act differently from default behavior--MORE SO than in a common blow-for-blow situation. This indicates it at least is programmed to do the equivalent of: observe, process, verify, reason, and execute. The first four are mental processes in learning.
Another example I have is playing the same character over-and-over. Yesterday, I played a Mario vs. MK battle several times. By the third match, MK was evading my approaches consistently for the first time that day. Why? Why was MK acting differently when I wasn't?
Either he had some ability to remember or his assigned initial behavior is random.
Is the cpu's initial behavior random? Maybe so. Even then, it still has the ability to observe patterns, process it, verify what it is, reason through its implications, and execute its most rational idea for behavior.
Also, everyone should know that the AI almost has a rubber-band nature to it. For instance, if I take a stock lead, regardless of my damage, the cpu player will play faster and smarter.
Why is this not so during the first stock? How/ why does its behavior change?
It adapts, presumably through a programmed ability to do so.
Because it can adapt, it can learn from situations.
Even if it is programmed to move in such a way when an opponent moves in such a way, its play is faster, smarter, and more intense. It is programmed to adapt; therefore it has the ability to learn situationally.
How about when I word it "they can adapt"?
I play against cpu players a LOT and one thing happens consistently: when you start to take control of the match ( i.e during a long, dragged-out combo), the cpu player will do everything it can to run away and camp--including MK. That indicates something exists making it aware it is in danger if it doesn't act differently from default behavior--MORE SO than in a common blow-for-blow situation. This indicates it at least is programmed to do the equivalent of: observe, process, verify, reason, and execute. The first four are mental processes in learning.
Another example I have is playing the same character over-and-over. Yesterday, I played a Mario vs. MK battle several times. By the third match, MK was evading my approaches consistently for the first time that day. Why? Why was MK acting differently when I wasn't?
Either he had some ability to remember or his assigned initial behavior is random.
Is the cpu's initial behavior random? Maybe so. Even then, it still has the ability to observe patterns, process it, verify what it is, reason through its implications, and execute its most rational idea for behavior.
Also, everyone should know that the AI almost has a rubber-band nature to it. For instance, if I take a stock lead, regardless of my damage, the cpu player will play faster and smarter.
Why is this not so during the first stock? How/ why does its behavior change?
It adapts, presumably through a programmed ability to do so.
Because it can adapt, it can learn from situations.
Even if it is programmed to move in such a way when an opponent moves in such a way, its play is faster, smarter, and more intense. It is programmed to adapt; therefore it has the ability to learn situationally.
kupo15
Smash Hero
But Corruptfate said that his AI Pit did a wingdash offensively against him. If they dont learn, how would you expect a CPU to know that:
-Jump>super quick Up B>left/right joystick movement>Dair cancel (wingdash)
would be effective for spacing over a basic movement? They would never think about doing this on their own
This has to be proof that the CPU must have learned by Corruptfate effectively performing this AT thus the CPU decided to store that button input and use it effectively.
I believe they do learn and if this really is the case, CPUs could possibly show us which ATs are actually useful. Ex. ppl think wingdashing is useless let the CPUs perfectly performed it to prove its usefulness.
This was just an example, please do not start a wingdashing usefulness discussion. It was not my intention.
Junahu
Smash Ace
I see nobody will be doing any serious testing to prove this one way or the other.
That makes me a sad panda.
That makes me a sad panda.
Shurk
Smash Cadet
I believe that mostly what we are seeing is an enhanced amount of circumstances being put into the code, Simply put "If, then " statements. Of course on a much more complicated level. So perhaps it is registering the button combinations you do, and deciding exactly how effective they are against you. Such as certain moves that people use a good amount of the time, I'll give it an arbitrary number of around 60%. A move you use successfully 60% of the time, may trigger a few lines of code to state something along the lines of "Watch out, counter with Airdodge", or "At --% do ---to avoid -- which happened 60% of the time"
Which is extremely plausible in a game such as this, therefor giving the effect that they are learning, when in fact the computer is merely compensating within it's own definitives.
Which is extremely plausible in a game such as this, therefor giving the effect that they are learning, when in fact the computer is merely compensating within it's own definitives.
goku21
Smash Rookie
MorphedChaos
Smash Lord
I think computers learn, as I've been CG'ed and infinited from 0% to death by a D3 computer level 9, and whenever I fight a MK computer, it just spams the tornado/Dsmash, and will occasionally do an IDC on me for like 20 seconds. (That sucked majorly...) I've also had Falco computers mini-hop laser me and CG me/Dair me off of FD before quite often, and my CPU ganondorf just loves to Ganocide me if he can get the chance... I hardly doubt htis is coincidence.
Frogsterking
Smash Journeyman
I don't think so. Someone mentioned crouch taunting earlier, and the computers did that to me once and I NEVER do that. We give them too much credit.
That was funny.
I think those could very easily be coincidences. Do you do those tactics with all of the characters yourself? And what stage were you playing D3 on? That might be part of it.
Testing would be interesting, and is the only way to clear this up.
CorruptFate
The Corrupted
That is the worst arguement for anything anyone could ever have given you should be removed from the gene pool for saying that.
Im not saying that they learn well or know what they are doing when they to it in some crazy terminator fan fic, but for those who say they dont learn please explain (i would love another expination) how/why my cpu's teabag and wingdash only after a player has done it an extesive amount.
