Best Brawl+ AI?

[Kei_Takaro]

I actually saved a video of me playing Link last week against a Luigi. He got off a pretty fancy combo on me that ended in a sweet-spotted UpB.

Somehow, he sweetspots me alot <_<;;

 

[MK26]

join the club

 

[JamesRaynor]

I'm no programmer, but I know for a fact that computers cannot be 100% random. In theory, the AI might "learn" by filling in data that you create (chaingrabbing, combos, etc) as a faux-random. I'm not saying learning from replays, but more of a "monkey see, monkey do" learning style.

Problem is... How do they know when to use or or how to incorporate that data if it works?

 

[Dantarion]

Computers can't be 100% random, thats true, but your theory is busted, unless you mean CPU's learning things during the span of one match, which could work, but is DUMB

 

[MaxThunder]

i feel a little ignored... anyways... going to test my skills against luigi... hope i do better against him than with him...

 

[GHNeko]

Computers can't be 100% random, thats true, but your theory is busted, unless you mean CPU's learning things during the span of one match, which could work, but is DUMB

...Who's game is this again?

lmao.

 

[MaxThunder]

nope... had no chance at all... it could have something to do with me not even being able to sit straight now... i'll try again when i'm not feeling so sick...

 

[IndigoFenix]

@Shadic Geno example.
Are you sure the CPU isn't just spamming the **** out of the A button?
I see no reason to believe that the CPU actually acts according to how the moveset is set up.

Just to point out, that wouldn't work. Geno Whirl only becomes a kill move if you press the A button ONLY right before attack hits the opponent. If you press it too soon, it becomes weaker instead.

I usually don't see the CPU land a timed hit, but then I don't use the Whirl too often, so if it is learning, that makes sense. If it isn't, the idea that the CPU can detect the moment to press the button by scanning the data in the attack itself is pretty impressive, though not impossible (it may well have been easier to program this than trying to input specific commands for every character in every situation). I find it hard to believe that the CPU would be able to identify such a precise moment in time through observation and imitation alone, though, and I highly doubt that SSBB has a more advanced AI than this 'string-learning' method.

I'll say again, useful techniques such as chaingrabs and edge-hogging are not indisputable evidence to the CPU's learning, since these are useful techniques and may very well be pre-programmed or a result of the CPU simply attacking where there's an opening. Some people say that if they use it more, the CPU uses it more, but then if you don't use it often you probably wouldn't recognize it. I'm more interested in things like crouch-taunting, a useless habit that could only be a result of either learning, or that it was actually pre-programmed as a sort of weird programmer's joke. Again, unlikely but not impossible.

The opening of the replay files seems to be good evidence to the idea that CPUs don't learn, unless the data is stored in an unexpected place. Has ALL the information in the replay folders been definitively accounted for? Can a replay file be opened, scanned, and every detail of its information accounted for with no undefined residue? Is it possible that AI data may be, say, inserted into the same area that the button-press data is contained - for whatever reason? Sounds inefficient and illogical, but then we don't really understand how the AI, if it does exist, works. It may be saved in the form of 'reactions' to events that took place in the match, for example (that is, a string of learned information may only be saved and 'inserted' into the CPU at the time it occurred during the match, which would help save on memory space since all of the learned behavior might not need to be recalled during a particular replay). Unless ALL replay data can be identified, accounted for, and translated, the ability to open up replay data and identify the names and purposes of files can only be considered evidence, not proof.

 

[Evilagram]

From the evidence I see, it would appear that every match, an algorithm generates a "seed" for the CPU's actions that is based on random data, and prior matches. A replay obviously has the same "seed" as the first time, so unless a code somehow affects the physics, or something like that, it'll replay the same as the first time.

The behaviors presented are very human in characteristic. I don't see where programmers would implement those behaviors, especially dash dancing in a game where dash dancing did not exist until we meddled, so it would appear to be learned by some simple heuristic.