AI's quality difference between SCIV and SC BD

sheppard

sheppard

[12] Conqueror
Have you guys noticed that in SC4, when you play against hard opponent, it uses cheat! How? Well I'll explain.

When you try to land an attack to hard AI opponent, it reacts so quickly. Every time when I try to land an attack, AI reacts with a faster attack to stop me if I'm in its reach. No human can do that. And like in Tekken 6, when you find a move which AI can't react or block, you can use that move to win almost all matches.

But in Broken Destiny, AI acts like a human. It rages and spams the same move if it has no other choice to win. AI uses different tactics and moves on same weapon users. It tries juggles on you and if you fail to block every time, it won't change it's tactic but when you learn how to block and get rid of it's juggles, it changes tactics.

I hope SCV AI will be like BD's AI. Or maybe better.
 
Yes, edgemaster sucks, hard. It has the perfect counter to almost every move. Also, if you start spamming the same move, it will start JIing it.
 
Yes AI is really good in recent SC serier. Even in SC4 its much much better than in previous games or Tekken.
 
Is it really better? It definitely seems more human but... I don't like SC4's A.I at all. The CPU seems really dumb, not only it doesn't finish combos, it doesn't seem to know any decent combos for each character , also doesn't break grabs, doesn't punish etc and gets hit by some move strings all the time etc. It's too easy to exploit its weaknesses therefore it has little value for someone who wants to gain some experience from it (apart from getting to know move animations maybe). I remember SC3 having some ridiculous A.I at highest difficulty level (was it "ultra hard"? I don't remember) - it was scripted to break 95% grabs, block everything, duck everything ... inhuman, yes, but also challenging and imo more fun.

BD has, without a doubt, the best A.I the SC has ever seen. But that doesn't mean it's perfect because it is not - the CPU likes to cheat here as well, for example doing Hilde's charged attacks numerous times in a row which is normally impossible.
 
Since I have absolutely no offline competition I pretty much live off playing Edge Master CPU (Ill take that over my shitty 4-Bar connection). I gotta say its much like playing a scrub who knows every single move animation/ has unbelievable reaction time. Mix-ups don't work because the computer doesn't hesitate, its just programed to attack like a beginner would.

I feel that the only way to play them is by playing with some basic rules similar to online:
I dont spam grabs, I never repeat moves which the computer seems blind too, I use exactly the same Mix-ups I would in real life even though I know they wont land against the CP, and I dont punish when he uses ridiculously unsafe moves that a real person wouldn't. Makes fighting them harder but at least I get to learn move animations and practice.

I agree with Ring though the CPU sucks. Its gotta learn some combos, play a safer game, punish, break grabs, and learn some mind-games. It honestly cant be that hard to program considering everyone says SCBD was close to perfect.
 
I haven't really had any problems with "cheating AI" in SC4, but AI in SC3... hooboy, that AI programming was just lazy (maybe they didn't have any calculation power left after all the "graffix"). It had insane reactions (it probably just read your inputs), but at the same time it ate stupid shit like there's no tomorrow (like Sophie's BB).
 
I liked SC3's A.I because it was so god-like and inhuman. Most of the time, it would duck even your fastest high attacks on reaction, break every grab, block every low etc, there was almost no room for mistakes, you had to take advantage of every mistake the CPU made to win. Still, I'd take a 4 bar online over SC4 CPU any day. Nothing like playing against human even though you can't react to many things...
 
Why don't they just take the best players from every character and model the AI after those guys? If modern technology can make a chess program play just like Bobby fischer or Garry Kasparov, then Namco can make the AI's IVY play like Malek or someone like that. Also, the AI needs to "learn" and "adapt" like a person would, while still sticking mostly to it's "preferred" moves, like a human would. Take away the stupidly fast reaction time that the computer has, and make it more based on an average human reaction time. If no human can GI a move that's 11 frames, then neither should the computer.

The computer should definitely be more combo oriented. It should also have the ability to "learn" new combos, by trial and error. Also the AI needs the understand mindgames, if move #1 is blocked 90% of the time, then try move #2, If move #3 is always punished, then do #3 less often.

This kind of AI might be unfeasible though, as it might be too much for the CPU. Also whatever the AI "learns" by playing, has to be a permanent adaptation. A simple game disc cannot "learn" new information because it cant store it. This means that the AI will never learn and adapt the way a human does, and unfortunately the ability to adapt is what separates the scrubs from the pros.

I don't think we should ever expect the AI to be as human-like as we want it to be.
 
damn-I-Suck said:
This kind of AI might be unfeasible though, as it might be too much for the CPU. Also whatever the AI "learns" by playing, has to be a permanent adaptation. A simple game disc cannot "learn" new information because it cant store it. This means that the AI will never learn and adapt the way a human does, and unfortunately the ability to adapt is what separates the scrubs from the pros.
Click to expand...

It is possible to store data as save data, I have seen examples but i can't remember right now. I think if they want human like AI, it'll need a hard work but it is possible.
 
damn-I-Suck said:
Why don't they just take the best players from every character and model the AI after those guys? If modern technology can make a chess program play just like Bobby fischer or Garry Kasparov, then Namco can make the AI's IVY play like Malek or someone like that. Also, the AI needs to "learn" and "adapt" like a person would, while still sticking mostly to it's "preferred" moves, like a human would. Take away the stupidly fast reaction time that the computer has, and make it more based on an average human reaction time. If no human can GI a move that's 11 frames, then neither should the computer.

The computer should definitely be more combo oriented. It should also have the ability to "learn" new combos, by trial and error. Also the AI needs the understand mindgames, if move #1 is blocked 90% of the time, then try move #2, If move #3 is always punished, then do #3 less often.

This kind of AI might be unfeasible though, as it might be too much for the CPU. Also whatever the AI "learns" by playing, has to be a permanent adaptation. A simple game disc cannot "learn" new information because it cant store it. This means that the AI will never learn and adapt the way a human does, and unfortunately the ability to adapt is what separates the scrubs from the pros.

I don't think we should ever expect the AI to be as human-like as we want it to be.
Click to expand...

It would certainly be possible to do this (the data just becomes part of your save game on your console's hard drive). I intend to create such an AI for the fighting game that I'm writing. Something like this should do the trick nicely: http://en.wikipedia.org/wiki/Reinforcement_learning

With an online connected game, it would even be possible to harvest play data from all games played, send it to Namco's servers, and train up an AI there that is the sum of all of the play experience of the game (and then offer that as an optional difficulty mode for people to play against). This is the fundamental flaw with any machine learning algorithm, they take a *lot* of sample data to learn anything useful, it would start out incredibly stupid (effectively purely random at first).

However, I suspect, that given enough data, it would just evolve (or devolve depending on your viewpoint) into being a frame-perfect AI with frame-perfect reflexes (maybe it would learn to just impact everything). It will observe things like i10 moves being guard impacted and incorporate that into its game as that was a good response (even if it was a total guess by the human player who first gave it that data point), or it'll observe two players at neutral inputting a i13 and i15 move simulataneously and learn to counter the i15 move on the first frame as that was a good outcome for that player. Its offense would become composed almost entirely of moves that work effectively against you, so, perhaps its best to just use a learning method for determining which attack to use (it'll learn to optimally punish and combo this way), and leave it up to the current method of randomly deciding whether it blocked/ducked/jumped/impacted/got-hit-by your attacks.
 
andur said:
However, I suspect, that given enough data, it would just evolve (or devolve depending on your viewpoint) into being a frame-perfect AI with frame-perfect reflexes (maybe it would learn to just impact everything)
Click to expand...

Cyberpunk Soul Calibur 5 :D
 
What drives me nuts about SC4 AI is that often they have actions that make little to no sense. I've given 5 examples below, keep in mind I'm not saying these are always the case, but it happens often enough.

Example 1, they sidestep an attack, then do nothing, no attack, no block, they just stand there and watch you attack that fly in your face or other invisible thing that your character is apparently so intent on slashing to pieces.

Example 2, excessive guarding, if you are really hating the AI you can play the, who can guard the longest without attacking game.

Example 3, SC4 was my first fighting game, and while learning it never ceased to amaze me that one could simply grab the AI to win, no other attacks necessary. This made normal story mode pathetically easy, even for a beginner. (There still should be some easy story mode for beginners, but normal was way too damn easy)

Example 4, the AI on the harder difficulties seemed far too polar. Either they were extremely defensive, or fairly aggressive. The defensive skill of the hard AI was (and still is) ridiculous, yet there offense lacks the relentlessness that human players use in their attacks.

Example 5, my favorite case of how bs the hard AI can be in SC4 is Algol on the top floor in the tower. I easily beat him in one try using yoda, and I am terrible with yoda. Why did I choose yoda? So that he couldn't just spam me with grabs while blocking all of my attacks.

Overall I think that SC4 AI are far too defensive and their amazing mind reading powers are a bit too frustrating at times. As a contrast to that, even the hardest AI don't attack enough, they tend to wait for you to attack then react, rather than taking their own initiative.

So yeah, there is a lot of room for improvement in SC4 AI.
 
You must log in or register to reply here.
Back