CPU's learn *video inside*

[saviorslegacy]

Okay... I know this has been posted before so I searched for it and looked on the first page of this forum. I didn't see it so I just decided to post a new one (either way there would be a new post).

Anyways.... there was a post saying that CPU's learn from human players. This match I had yesterday proves that.
http://www.youtube.com/watch?v=HlI5PnGtFuc&feature=channel_page

 

[Project D]

Spoiler

rofl@ the dash dancing part

lol, I'm amazed CPUs could do some human things.

 

[saviorslegacy]

rofl@ the dash dancing part
lol, I'm amazed CPUs can do so much human things.

*fixed

I know... I was shocked! I was first fighting him like he was a CPU, but I had to change my style in mid battle (plus he was doing things that made me go "WT!").

I wonder what Sheik will play like in a year?
At least I know where to fight myself.

 

[Hiiiiiii]

Maybe we should just start putting our CPUs in tourneys instead of us

 

[ThreeSided]

I saw that, I was like, "dude, cool. Doesn't quite prove it... could just be coincidence."

Then it did the dash dance. x_x

This has to irrefutably prove that CPUs learn.

You have no ****ing idea how hard I laughed when I saw that dash dance.

 

[ndayday]

I'll go ahead and be skeptical and say it's a code to make a human player's tag say CPU.

Nontheless, that is pretty cool.

 

[saviorslegacy]

I'll go ahead and be skeptical and say it's a code to make a human player's tag say CPU.

Nontheless, that is pretty cool.

You could if you wanted too.... I don't really have proof though. However... I was talking to <3 on Skype while I was doing it and I was going "WTF!". So he can at least say that he was talking to a suprised me. Oh... and my bro saw it too. So go ahead and believe what you want.... but that is a CPU.

I agree... really cool.

At theesided: My CPU Sheik does stuff like this all of the time. It is just that this one dash danced. So I just HAD to save and upload it.



BTW... I am depressed that the CPU use's my play style better than I do.

 

[ThreeSided]

Hm... watching again I am getting a bit more skeptical. The CPU isn't shielding nearly as much as CPUs usually do.

 

[saviorslegacy]

Hm... watching again I am getting a bit more skeptical. The CPU isn't shielding nearly as much as CPUs usually do.

*the CPU learns from the humans
I didn't sheild like crazy.. so why would he?

Also note how what I gimped myself he done to USmash's for no reason. There was also a few other things that a human player just wouldn't do that he did. (like rapid jabs so much)

I DO NOT hack (ask almost anyone that knows me) and this is not a hoax.

 

[kr3wman]

OOOOOOOOOOOOOOOOOooooooooolllllllllllllllld.

Maybe you'll use the search button next time.

 

[<π]

SL didn't hack. I was talking with him while he was playing this match

 

[saviorslegacy]

OOOOOOOOOOOOOOOOOooooooooolllllllllllllllld.

Maybe you'll use the search button next time.

WTF I said I searched. I didn't dig to deap though.
BTW... has anyone ever posted a vid of a CPU doing a dash dance before? If the answer is no then this is brand new and proof that will end the arguements over "CPU's learn".


:EDIT: Thank you <3.

 

[Cacti]

That CPU seemed a lot better then most I've seen. Until the dash dance, that's all I thought. I went "WTF?!?"

 

[MajorasMask999]

If CPUs learned over time, then wouldn't that mess up replay data? Because the replay data is just the human players' button inputs, and the conditions set before the match was started (characters, colors etc.), right? If CPUs actually learned over time, then replay data wouldn't work, because they would be acting different in relation to your actions than they did when you first did the match.

I've also seen a CPU dash dance before, but it looked like it wasn't doing anything strategic. It looked like it was just some random coincidence.

 

[Cacti]

Why would it affect replays? Remember that it just records the CPU's inputs, and there's not an actual CPU in control.

 

[MajorasMask999]

Why would it affect replays? Remember that it just records the CPU's inputs, and there's not an actual CPU in control.

No, it doesn't record the CPUs inputs. Only the human inputs and the environment changes are recorded.

Like, one time I used a hack that sped up the level and saved a replay of it. When I watched the replay of it with the code still on, everything played back exactly as when I originally played it. But after taking the code off, the stage was going at it's normal speed. Since my inputs were recorded, I started going all over the place and ended up suiciding a bunch of times. The CPUs, on the other hand, remained on the stage and continued playing normally.

 

[bobson]

Why would it affect replays? Remember that it just records the CPU's inputs, and there's not an actual CPU in control.

No it doesn't. Take a replay made with gameplay-affecting codes on, then play it without the codes. Your inputs are totally out of whack, but the CPU plays exactly like it would normally.

If CPUs learned over time, then wouldn't that mess up replay data? Because the replay data is just the human players' button inputs, and the conditions set before the match was started (characters, colors etc.), right? If CPUs actually learned over time, then replay data wouldn't work, because they would be acting different in relation to your actions than they did when you first did the match.

Unless the replay data stores the CPU's configuration.

 

[WastingPenguins]

If CPUs learned over time, then wouldn't that mess up replay data? Because the replay data is just the human players' button inputs, and the conditions set before the match was started (characters, colors etc.), right? If CPUs actually learned over time, then replay data wouldn't work, because they would be acting different in relation to your actions than they did when you first did the match.

This is incorrect. Replay data saves the CPU actions as well as player button inputs. CPUs don't follow a static path exactly corresponding to the human player's actions. There is definitely a level of unpredictability programmed into their actions-- they have many different options at a given point and might perform any of them.

Think about it-- if it really worked like you say, then starting a match and simply standing still would produce the same exact CPU reaction, every time. It definitely doesn't.

 

[bobson]

Think about it-- if it really worked like you say, then starting a match and simply standing still would produce the same exact CPU reaction, every time. It definitely doesn't.

That doesn't happen because of the random seed the AI uses. The replay saves the particular random seed that was used for the match so the CPU's actions are aligned.
Like I said, play a replay made with codes without any codes activated and the CPU will act perfectly normal, even taunting when you die.

 

[Blank Mauser]

I've only had a CPU dashdance on me once. Middle platform on Lylat, it seems they like doing it on platforms. Don't know if its really a learned tactic or not.

 

[pure_awesome]

If the replay didn't record the CPUs actions at all, how would it know when it was supposed to trip?

It has to record some stuff.



That being said, CPUs don't learn and all the stuff in this video has been happening since the game came out.

 

[Iwan]

haha...its also been said by HAL laboratories that the CPU can actually read your controller input as you're doing it....

Aka they know what you're going to do and counter accordingly/that's why they powershield all the time.

 

[bobson]

If the replay didn't record the CPUs actions at all, how would it know when it was supposed to trip?

That would be covered under the random seed, wouldn't it?

 

[MajorasMask999]

Unless the replay data stores the CPU's configuration.

Would it really be able to store the "CPU's configuration" in a replay file ?

If it was just something small like a number (CPU Level), then that could be saved, but if the CPUs started to learn over time, it would have to store larger amounts of information in order to pull off the more "human" maneuvers, such as how to do what, and when and where to do it. It just doesn't seem to me like the kind of thing that would be possible.

 

[pure_awesome]

That would be covered under the random seed, wouldn't it?

...no?

Tripping is randomly generated any time the character smashes the stick to the side while in a "Tripping Possible" state. It wouldn't have anything to do with which particular seed the CPU was using at the time. It's independant of that. Same with G&W's Hammer, or what Peach pulls out of the ground.

 

[bobson]

If it was just something small like a number (CPU Level), then that could be saved, but if the CPUs started to learn over time, it would have to store larger amounts of information in order to pull off the more "human" maneuvers, such as how to do what, and when and where to do it. It just doesn't seem to me like the kind of thing that would be possible.

Depends on how many "learned behaviors" it's capable of storing and how much space they take up.

Tripping is randomly generated any time the character smashes the stick to the side while in a "Tripping Possible" state. It wouldn't have anything to do with which particular seed the CPU was using at the time. It's independant of that. Same with G&W's Hammer, or what Peach pulls out of the ground.

I was under the impression that the random seed was the base for determining other random aspects of the match. EG: Judgement randomness determines that there will be a nine 2 * x turns ahead, but the random seed determines x.
If it were true that random aspects were just recorded in replays, a CPU would experience the same randomness in a desynced replay, which I'm pretty sure they don't.

 

[pure_awesome]

You seem to know alot more about this than I do, to be honest.

If what you say is accurate (and you seem confident enough that it probably is), then that would certainly explain how it doesn't seem to record anything, but still manages to have a "memory" of random events.

Where'd you learn about the seeding stuff? I'd like to read up on it.

Also, I'm fairly sure you're right about the desynched replay not experiencing the same randomness, I've experienced it before with the frame codes.

 

[Prawn]

This is really interesting, hopefully someone with a bigger rep and better understanding of the mechanics of the game will come in here and give us a summary. Only thing that I could dispute with is that the dashdance may have been the computer making decisions and simultaneously canceling them in lieu of a better option. But that doesn't explain a lot of the other stuff.

Good show!
or, Good troll! Whatever one ends up being true

 

[mofo_]

No, it doesn't record the CPUs inputs. Only the human inputs and the environment changes are recorded.

Like, one time I used a hack that sped up the level and saved a replay of it. When I watched the replay of it with the code still on, everything played back exactly as when I originally played it. But after taking the code off, the stage was going at it's normal speed. Since my inputs were recorded, I started going all over the place and ended up suiciding a bunch of times. The CPUs, on the other hand, remained on the stage and continued playing normally.

same thing with me. i saved a replay of brawl+ and when i watched it was all good, i killed the cpu 3 stocked. when i loaded no hacks and watched it........i suicided many times ,but the cpu didnt suicide, and when i suicided it taunted! i was like " hey what the - "

 

[Wolf of Ice]

Maybe the data from it doesn't make sense to it without the hacks?

Is it possible the game recognizes something's not possible?

 

[bobson]

You seem to know alot more about this than I do, to be honest.

I probably don't; I'm just good at appearing knowledgable.

Where'd you learn about the seeding stuff? I'd like to read up on it.

Just general design knowledge. Randomness has to come from somewhere, and a base point is an easy way to derive it all. I have no clue whether or not Brawl actually uses the system, it just seems to work well with the evidence.

 

[sMexy-Blu]

I believe that the CPU wants to play Melee.

 

[Project D]

same thing with me. i saved a replay of brawl+ and when i watched it was all good, i killed the cpu 3 stocked. when i loaded no hacks and watched it........i suicided many times ,but the cpu didnt suicide, and when i suicided it taunted! i was like " hey what the - "

so if the CPU taunted and learns and sort of copy what humans do...
do you reckon that if you have a level 1 or really low level CPU, put a noob that taunts all day on like the biggest map possible, human player taunts all the time, and the CPU taunts as well?
possible theory lol.

 

[Xtreme Starfox]

... -.-

If CPUs learned, online would Desync, + Replays would go out of synch.

Replays do not save CPU controls, only player moves get saved.

 

[Sosuke]

CPUs copy button inputs that prove successful in matches between players.
Or something like that.

 

[Browny]

Nothing out of the ordinary whatsoever

Remember, the AI was programmed BY HUMANS in the first place, its not impossible for them to use techniques that the developers thought were good.

If i saw a single ftilt lock, jab-grab, dacus, or fall-through-platform aerial, then I might begin to believe this. So far its just a coincidence, and you searching for any and all similarities between your style and the CPU's, which is BOUND to happen in any match.

 

[TKD+ITA+Mar=]

My CPUs sometimes rapidly crouch.

 

[saviorslegacy]

My CPUs sometimes rapidly crouch.

That is them loling at you. -_-

 

[Drakar]

lol i was surprised up until the dash dance

then i was shocked

 

[Zap94]

I think that you may be right. I've noticed that when I fight computers that are my main or my brothers' mains, they use similar tactics to the ones that we use.