What is the best AI there is? Or what would can you dream of to make AI advanced?

[Flux]

Whats the best AI there is?
What components would make AI human-like and "learn" weither from constantly updated patches and human player's resources vs Ai resources efficiency statistics etc. What would you do(given you had the budget) to make the best most advanced AI?

Personally I know I am a minority nowadays with Nintendo's casual/non-intimidating approach dominating the video game console market but I want insanely hard/strategic/intelligently multi-semi-random options in AI. I want to gameplan. I hate easy games.

 

[eastmen]

IF you make it to good and where it can learn from playing against you then you will never be able to beat it. Whats more , if they go through a spore like database where the computer takes the infromation from everyones play through most casual gamers wont get past the first stage or two.

I like an a.i that will put up an interesting battle with me. It will duck for cover , call for back up and what not , but if I watch it and i play smartly I can beat.

 

[Nesh]

If we want a human like AI shouldnt we also factor the human error?

I can beat humans when I play with them. They arent superhumans. What makes humans more interesting opponents in games isnt just their level of skill. Humans are less predictable and you can find ways to manipulate them because humans have psychological fluctuations.

Whereas AI in games just reads your input and immediately adapts. Which is boring even if it is challenging

 

[Freak'n Big Panda]

If we want a human like AI shouldnt we also factor the human error?

I can beat humans when I play with them. They arent superhumans. What makes humans more interesting opponents in games isnt just their level of skill. Humans are less predictable and you can find ways to manipulate them because humans have psychological fluctuations.

Whereas AI in games just reads your input and immediately adapts. Which is boring even if it is challenging

I wonder how you'd go about simulating psychological fluctuations. Cause if you break it down that's the only difference between the human decision making process and input -> output AI we see today in games.

 

[assen]

You don't want an AI which you can't beat. In many types of games - shooters, RTSes, racing games, it's easier to make such an AI. What you actually want is an AI that fights tooth and nail to lose convincingly.

 

[Rangers]

I believe AI is a dead end street more or less. Carmack has some comments on such.

And if you want to make a game harder, "tougher" AI isnt really needed, just give the enemy more hit points, stronger weapons, etc. You can make the game arbitrarily difficult with ease.

Aggressive AI can provide a lot f fun, and is one of the high points of a game like Gears to me. However overall I feel like it's a very overrated part of game development. The bottom line is there's only so far it can go. I dont believe a Turing Machine can ever exist. In that vein, I dont believe AI can proceed beyond a certain point, and thats born out in gaming continually imo.

 

[tongue_of_colicab]

And more realistic AI wouldnt always make a game harder/better. Just look at me, I suck ass at playing games. If you'd made a AI like me even a monkey could beat the game no problem

 

[Acert93]

Not much time... but best "AI" opponents would be ... humans. Create a "SP World" where people online can enter the game as baddies. They may be weaker, slower, and be at a general disadvantage but give them some incentive to "play the role" of the bad guy and a lot of respawning and ta da great "AI"

In general though great AI serves the purpose of the game itself. In most games you are a super hero with the odds against you so the AI needs to be adapted in ways to make the game fun, challenging (but not too much), and set the tone for the events unfolding. So technically a game can have "weak" or "dumb" AI but they do a great job with placement, number, difficulty, and setting that reinforces the game design, which makes it "good" AI.

 

[Iron Tiger]

Totally realistic AI isn't necessarily fun. If the opponent is responding as an actual human would, in most cases, it would put the player at an unfair advantage. A real person avoids being shot because it limits mobility, and it HURTS. In games, we don't have the penalty of pain to motivate us, so we can run into a hail of bullets if we see a med kit at the end of the corridor, or a pocket to hide in and regenerate health. That would be a cool setup if we're supposed to be superhuman and the enemy is supposed to be terrified of us (like the covenant calling Master Chief "the Demon"), but on a WWII battlefield, it's terribly unrealistic.

If AI is going to be more realistic, our player's responses to damage have to be made more realistic, as well. Shoot us in the leg and we should stumble and then limp thereafter. Regeneration shouldn't be automatic or after a magical health pickup. Severity of wounds should matter, and blood loss should have an effect on stamina, dexterity, and balance - gradually depleting over time without treatment. It may seem unfair that you can be randomly hit on any part of your body and have a different effect, but that's reality. In an HD era, we should have to look at the opponent's weapon from across the field and make a judgement on whether it can penetrate our armor.

 

[Arwin]

AI is going to stay important I think ... whether it is to make sure thousands of NPCs behave suitably in the right context, or whether AI can lean to deal with new physics based gameplay, etc. Just pick off a list of all the best games that will be released this Christmas, and tell me in how many of them AI is completely irrelevant in all aspects of the game ... You're going to have trouble even naming one.

However, for competition purposes certainly human opponents are a big help. The big advantage of human opponents is that there is a huge range of them in terms of skill and predictability, and their AI is very adaptive.

 

[infinity4]

AI will advance, but will never be as good as human brain, because the way they work is different to each other. The main limitation is that brain makes millions of new synapses all the time, AI is limited in that sense.

 

[inefficient]

I think in general "AI" is probably the most difficult aspect of games for non developers to have a useful discussion about. Because it's all just smoke and mirrors behind the scenes. Even with the AAA games. The "best" games are just the ones that fool you the best.

When it comes down to it, there is no such thing as "Artificial Intelligence" at all. It's a total misnomer. It's just the illusion of intelligence. The games people usually point out as with "bad" AI are just the ones where the smoke and mirrors don't stay fully hidden and out of sight.

Convincing "AI" behaviors is not a matter of running more or smarter "AI code" on the CPU. It's not something you could improve linearly with more CPU power. It artists doing many more animations, a well designed world that masks and hides event triggers from you, a good world collision system, even things like more dialog so NPCs don't repeat themselves over and over breaking the illusion. It's the sum of so many factors. Most of which are totally out of the "AI programmers" control.

 

[Gradthrawn]

Convincing "AI" behaviors is not a matter of running more or smarter "AI code" on the CPU. It's not something you could improve linearly with more CPU power. It artists doing many more animations, a well designed world that masks and hides event triggers from you, a good world collision system, even things like more dialog so NPCs don't repeat themselves over and over breaking the illusion. It's the sum of so many factors. Most of which are totally out of the "AI programmers" control.

Bethesda on that very point:

10) How advanced will the AI of NPC's be this time around? Are they really going to have a life? Speaking to other NPC's in a logical manner, traveling and trading with/in faraway places, Submitting to the player rather than fighting if they know, or think, they're no match for him?

I wish I could answer with a number, like "it will be 17 advanced." AI is difficult to define, the NPCs certainly appear much smarter than our previous stuff, by a lot. Much of that is us giving them better data, massaging what they do so the player gets to see more of it. We added a lot of animations, so people in town are doing more. They "seem" to be interacting with the world in a more realistic manner, but that usually means going up to something and playing an animation. It can be something really simple, like we added "lean against wall". It's funny how something that small can give life to a person. They walk into a space, and just lean against the wall, arms folded. Like Oblivion, we use our Radiant AI system, so most of the NPCs eat, sleep, work, etc. I think we take it for granted now, but it's pretty great to have that level of control. We've also done a lot to the conversation system, which makes them seem a lot smarter, but again, that's better data, not a new system.

On the technical side we spent most of our time doing an all new pathfinding system. Morrowind/Oblivion use nodes for pathing and Fallout uses a navmesh. This is the difference between an NPC having a valid point to stand on (node) versus an area to stand in, or walk around (mesh). You can do much more sophisticated actor movement and behavior with a navmesh, and I think you'll see the results onscreen, especially when the bullets start flying. The actors do a great job of finding cover and using the space well, something we could never have done with pathnodes.

In terms of the NPCs traveling around, many travel around town, and some travel the wasteland. There are a few caravans in the game that go from town to town trading. Radiant AI handles something like that really well.

Lastly, as far as submitting to a more powerful foe, yes they do that, in that they run away. If they're overmatched, they holster their weapon, flee and try to hide. While this sounds cool on paper, it's often not fun at all, and we've ended up really dialing that back, because it gets really annoying really fast, to have people run away all the time. The main faction that still acts like this are the Raiders, the others don't do it so much.

I just happened to be reading the Fan Interview over there at the same time as this post.

 

[Flux]

For the record you can make human-like AI for a video game. BUT....

1)I agree with one of the posters about how a person would respond.

2)nothing comes close to a a human brain( pattern recognition, dynamic context of bevahior,variation,trial/error, pain/ pleasure responses,memory recall, sensory matrixed memory(sight memory linked to smell memory linked to taste memory).

3)The only thing I think would give you the closest thing to human-like AI is a few things...

an array of 1024 ZISC(1024 neurons each) cores that is dedicated to pattern recognition,memory recall,variation,simulated emotions,etc. and an VERY large FPGA that was designed to simulate a humans psychological responses(hardwired responses,subconsious,and memory recall, emotions, social recognition etc)

That alone would be expensive and somewhat useless for games...

 

[liolio]

I think that Bungies got it right for game the next step in AI is animation.

 

[psorcerer]

It seems like a lot of people do not understand that AI (in the computer science meaning) and so called "game AI" have nothing in common. Game AI doesn't solve any real-world problems, it's there just to entertain the player. In the most broad sense the game AI - it is the game itself, so the ultimate goal is indeed to lose in the most satisfying, entertaining and convincing way. To lose.
It's not a person who plays the game it's a game who plays with person. Too bad that even in game industry only one person seems to understand this concept completely: Hideo Kojima.

 

[BoardBonobo]

GaameAI shares alot of the methodologies that the science version uses, such as search paths, neural networks, some genetic algorithms, black box approaches, etc.

Some areas of gameai are pretty much sown up e.g path navigating and to some extent active scenario reactions. But language parsing and verbal responses could be more realistic that would be a good way too advance IMO.

 

[obonicus]

GaameAI shares alot of the methodologies that the science version uses, such as search paths, neural networks, some genetic algorithms, black box approaches, etc.

Some areas of gameai are pretty much sown up e.g path navigating and to some extent active scenario reactions. But language parsing and verbal responses could be more realistic that would be a good way too advance IMO.

Game AI is mostly graph pathfinding. Games that actually use neural networks are really really rare (are there any commercial games? I know of a few research projects that say). Natural Motion/Euphoria (I forget which is the company and which is the tech) apparently uses a neural net to tweak the animations, but it doesn't do it in real time. Science AI is... well, it doesn't even use the word AI anymore, because the word now has an unfavorable connotation, so instead you see terms like 'machine learning'. The big techniques are mostly probabilistic and I'm not even sure how useful they'd be in a game, even if they could be run in real-time.

 

[ban25]

I think in general "AI" is probably the most difficult aspect of games for non developers to have a useful discussion about. Because it's all just smoke and mirrors behind the scenes. Even with the AAA games. The "best" games are just the ones that fool you the best.

When it comes down to it, there is no such thing as "Artificial Intelligence" at all. It's a total misnomer. It's just the illusion of intelligence. The games people usually point out as with "bad" AI are just the ones where the smoke and mirrors don't stay fully hidden and out of sight.

Convincing "AI" behaviors is not a matter of running more or smarter "AI code" on the CPU. It's not something you could improve linearly with more CPU power. It artists doing many more animations, a well designed world that masks and hides event triggers from you, a good world collision system, even things like more dialog so NPCs don't repeat themselves over and over breaking the illusion. It's the sum of so many factors. Most of which are totally out of the "AI programmers" control.

If there was still a recommendation system on this forum, I would give a +1 for this post. Most game AIs are just state machines with some simple decision trees. What players usually perceive as 'better AI' is usually just a series of more elaborate interactions (animations, to be specific).

A lot of the tuning in action games usually involves some degree of dumbing down the enemy so it's not overly difficult or frustrating, usually the latter. One game I did some AI work on had an enemy that would flee and circle when attacked. This was extremely effective in terms of harassing the player, but it was also very frustrating because every time you hit the enemy, he would run away. Throw a group of these at a player and he's liable to chuck his controller in the trash. In any case, it was tweaked such that the enemy only had a 25% probability of fleeing when attacked (or something in that range).

 

[Flux]

Game AI is mostly graph pathfinding. Games that actually use neural networks are really really rare (are there any commercial games? I know of a few research projects that say). Natural Motion/Euphoria (I forget which is the company and which is the tech) apparently uses a neural net to tweak the animations, but it doesn't do it in real time. Science AI is... well, it doesn't even use the word AI anymore, because the word now has an unfavorable connotation, so instead you see terms like 'machine learning'. The big techniques are mostly probabilistic and I'm not even sure how useful they'd be in a game, even if they could be run in real-time.

Good points obonicus. I think Game AI should be made with a model like sports. Its not fun to play a opponent that you "know" will lose to you. But it is fun playing an opponent that will lose to you if you outsmart/outperform its defensive assumptions. Any and all games boil down to PSR(paper scissor rock-like guess/countering games) and defensive/offensive assumptions based on "the tape" or historic patterns(patterns you do over the duration of playing the game)and timing. The game can just run statistics on what you will likely do to...

-Counter it(offensively ie does take damage or resources away from the CPU)
-Defend it (non-offensive ie doesn't take damage or resources away from the CPU)

All the programmer does is give it a laundry list of tricks/traps(defensive/offensive schemes) that the game program references so they can find what better beats you. The CPU also finds does subtle ways to throw your timing off by alter the...

-Angles
-Timing(offensive attacks)
-spawn spots(some can be outlawed by the programmer because this can be broken)
-Resource placement/spawn rate(see above about limitations)
-Feints(potentially dominant tactic actually)

The AI can also be programmed with a library of intelligent hardwared team offensive/defense schemes. These schemes can be tweaked by the AI on the fly to throw off the human player.

-Selective flooding(flood the player with enemies if they are low on health or have non-crowd control weapons then when they adapt change the attack to another scheme)
-Runaway games(like
a)when low on health run away
b)hit and runs
c)draw into ambushes
d) steal items and run
e)snipe and run... etc)
-Playing stupid(draws)
-Kamekaze enemies(hey blow me up with your high tier weapon and use up your resourses to soften you up for the real threat)
-Delusion draws( ok its a item/hostage/tool etc....what an enemie halogram?)
-Some enemies can heal themselves off spawned hp items(potentially broken)
-The Okay dokey (I think the enemie can't go through walls...whoa! it went through a wall! Ok its a shotgun....Whoa! its a rigged pipebomb! 4-5 seconds to leave to room!)
-Randomizing (well I know if I go through room A and carry X weapon then enemie Y will apear on the north west corridor.....What an enemie z on the ceiling with a cloak trap and the enemie in the north west corridor is a enemie pretending to be a hostage so I can't hit them with a weapon I have to go up close to the hostage and make sure that the hostage is a person AFTER being ambush by ceiling enemy and softing me up for the fake hostage. The health pick ups both spawn on random time intervals AND spawn in random spots(a set percentrage of the spawns are advantagous to the CPU aka spawn inside a trap or spawn right next to an enemie. Oh an semi-random enemie stats.)
-Semi randomly generated stages
-Sabatage(dudded items, trap quests,fake hostages,sabatoged weapons, enemies that are hostages(cloaked by the CPU to get you to kill them), hostages that are suicidal,hostile hostages (they fight you and you have to convince them to stop through non lethal means....oh and enemies are attacking you both...good luck).
-Enemies that are simulated to be the same as you...(using the same attack patterns as you.They have powerful weapons like you do AND they can heal,powerup or call for help).
-Enemies(that don't look like enemies) that predend to be tame and benevolent and help you on your quests turn on you(in a random way not all the time) when your in some some crazy bad situation. ie they lock the door when your in a room looking for pistol rounds then a pheromone capsole is rigged to break. Enemies flood the room as you find a way out. Yes there is a way out. The Pretending Hostage then tells you to give them all your attack weapons or they won't let you out.You then spend some time trying to get the trust of your new captor have fun your now a hostage!


Game AI should be about "how to be a human palyer being the CPU 101". Using timing,PSR with statistics,mind games,etc you keep the human player from just memorizing or planning thier way around the game(boring and repeatitive).Why do you have this level of AI in the game?

Replay value