"Future Gen" Ai in video games.
First things first, I think this could benefit from being in the AI research forum. BUt this is just an opinion.
As for the system, I believe that to build up an individual database for each MOB and NPC in a game, you would need an unimaginable amount of space to store everything, a fair part of which would be blank space to allow for more info to be added, and which would be alternately cleaned. I don't think this is the way to go. BUt I can't say I have honestly looked into the subject very much. TO be honest, it sounds a little like PnP Rolemaster tables to me. You store ALL the info anyone can think about in books, and then choose which pages said characters can have access to, or what rules you're going to forgo, and you constantly adapt your way of playing. It is okay when you have a GM doing the thing, because he can choose to adapt or not. But playing Rolemaster with an unimaginative GM is really a pain, because you almost have to roll dices to know wether or not you can open a door, or unsheath your weapon, without hurting yourself. Sometimes, it gets quite boring having to wait for the GM to shuffle through his books. I'll grant you that the computer may shuffle faster, but it would still require a fair amount of computing, and of Disk work, which mean both heat, noise and slow playing if not at top conditions, which combined give the player a fair share of headache... Try and find something else.
Of course, you all feel free to comment uppon what I just said, even to destroy it completely...
As for the system, I believe that to build up an individual database for each MOB and NPC in a game, you would need an unimaginable amount of space to store everything, a fair part of which would be blank space to allow for more info to be added, and which would be alternately cleaned. I don't think this is the way to go. BUt I can't say I have honestly looked into the subject very much. TO be honest, it sounds a little like PnP Rolemaster tables to me. You store ALL the info anyone can think about in books, and then choose which pages said characters can have access to, or what rules you're going to forgo, and you constantly adapt your way of playing. It is okay when you have a GM doing the thing, because he can choose to adapt or not. But playing Rolemaster with an unimaginative GM is really a pain, because you almost have to roll dices to know wether or not you can open a door, or unsheath your weapon, without hurting yourself. Sometimes, it gets quite boring having to wait for the GM to shuffle through his books. I'll grant you that the computer may shuffle faster, but it would still require a fair amount of computing, and of Disk work, which mean both heat, noise and slow playing if not at top conditions, which combined give the player a fair share of headache... Try and find something else.
Of course, you all feel free to comment uppon what I just said, even to destroy it completely...
If I'm reading your idea correctly I have a similar idea that I haven't put into practice yet. I have done some theoretical overall design for it though. That idea is based around a computer simulation of memetics. All agents in the world have a set of meme that comprise their knowledge of the world, and then information is spread between agents through conversations and similar means.
I'm still planning to try to implement a basic test of this someday, but real life has come in the way so far...
I'm still planning to try to implement a basic test of this someday, but real life has come in the way so far...
I suggest moving this to the AI forum too.
Also, you were talking about having a data-base full of AI? And how big are the data-base going to be?
It is pretty easy to tell a computer to learn, but... "It is impossible to tell a robot what NOT to learn"
Also, you were talking about having a data-base full of AI? And how big are the data-base going to be?
It is pretty easy to tell a computer to learn, but... "It is impossible to tell a robot what NOT to learn"
I think what he means is that he wants a database of knowledge/concepts that AIs contribute to, and this information is redistributed to other AIs (based on things like their location in-game, their history, and who they've come in contact with) so they don't have to do the thinking all over again. This would allow more intelligent AIs than would be possible with completely independent ones.
There is a problem though.
Lets say AI finds a good solution and wins - now AI knows that good solution and uses it.
Now lets say AI failed and lost. It then knows that this move is bad and looks for another one. Fails again. And again. The problem is that if AI is unable to find solution fast it wont find it ever. AI operating by some algorithm can't possibly understand why it failed. Such that it can't look for new solution even if it knows all previous outcomes or experiences of AI at the same situation.
Lets say its a FPS. An AI operates a soldier in some room. For isntance there is a spot to hide and ambush a player. If AI won, it will always be there at a good spot - no problem. But lets say it tried and lost. Why?
- Was it because player spotted legs of prone AI soldier? Maybe standing will totaly cover AI soldier at this spot, but AI does not know it and can't comprehend.
- Was it because player had luck and tossed a frag behind some boxes without any clue? maybe, but AI does not know.
- Was it a shade from fire which accidentely was lit by player fire? Could be, but AI does not understand that.
All it does understand is - my previous experience with hidding here was a loss, I wont do it again, I need new solution. And AI stands in the middle of a room...
The problem is the AI can't really understand any data in any creative way and analyze it. Human will cry "omg how?!!!" and then think out of the box. AI will just record sequence of movements and outcome.
PS
The problem will also be that AI will do same thing over and over again if it had good outcome. So it will lose element of surprise.
The most amazing thing about human intelligence in a game is that you never know how stupid it could be, such that sometimes you just discard some valid moves from your analysis because they are stupid, yet enemy does exactly that.
Lets say AI finds a good solution and wins - now AI knows that good solution and uses it.
Now lets say AI failed and lost. It then knows that this move is bad and looks for another one. Fails again. And again. The problem is that if AI is unable to find solution fast it wont find it ever. AI operating by some algorithm can't possibly understand why it failed. Such that it can't look for new solution even if it knows all previous outcomes or experiences of AI at the same situation.
Lets say its a FPS. An AI operates a soldier in some room. For isntance there is a spot to hide and ambush a player. If AI won, it will always be there at a good spot - no problem. But lets say it tried and lost. Why?
- Was it because player spotted legs of prone AI soldier? Maybe standing will totaly cover AI soldier at this spot, but AI does not know it and can't comprehend.
- Was it because player had luck and tossed a frag behind some boxes without any clue? maybe, but AI does not know.
- Was it a shade from fire which accidentely was lit by player fire? Could be, but AI does not understand that.
All it does understand is - my previous experience with hidding here was a loss, I wont do it again, I need new solution. And AI stands in the middle of a room...
The problem is the AI can't really understand any data in any creative way and analyze it. Human will cry "omg how?!!!" and then think out of the box. AI will just record sequence of movements and outcome.
PS
The problem will also be that AI will do same thing over and over again if it had good outcome. So it will lose element of surprise.
The most amazing thing about human intelligence in a game is that you never know how stupid it could be, such that sometimes you just discard some valid moves from your analysis because they are stupid, yet enemy does exactly that.
Feedback influencing behavior is at least as old as Wonder Project J and Galapagos.
I remember Killer Instinct for the SNES figuring out my favorite combos and becoming more and more likely to counter them, so that by the time I reached the final boss I couldn't land a single hit on that bugger and had to switch characters and change up my style to win.
I think it can be done.
Put a matrix of properties and values under the game's taxonomy. A sword would have component values for weapon, edged, steel, etc. A character who has a very low affinity for weapons would flinch at the sight of a sword, whereas a metal-loving character might be fascinated by it. Shared memes would probably be appropriate for representing shared knowledge, like the player's fame and standing with a given faction or race. So if the Orcs have a low standing with you and you're wielding a weapon that's got an anti-orc property they can recognize and your armor is made by druids, the enemy of the orcs, then you'll be extremely unpopular with the orcs. If you walk into a town wearing inappropriate clothes or pay with the currency of an enemy faction, that'll be automatically factored into NPC reactions. First impressions would count for a lot, but if you do enough work to get your standings up, your clothes and hairstyle will have less impact in the face of a more fleshed-out "profile" with that faction or individual.
I think the data needed for such interactions could be compressed pretty efficiently, and would never come close to the vast space occupied by texture files in modern games. How much space is occupied by all the relationship data in a Sims game?
I remember Killer Instinct for the SNES figuring out my favorite combos and becoming more and more likely to counter them, so that by the time I reached the final boss I couldn't land a single hit on that bugger and had to switch characters and change up my style to win.
I think it can be done.
Put a matrix of properties and values under the game's taxonomy. A sword would have component values for weapon, edged, steel, etc. A character who has a very low affinity for weapons would flinch at the sight of a sword, whereas a metal-loving character might be fascinated by it. Shared memes would probably be appropriate for representing shared knowledge, like the player's fame and standing with a given faction or race. So if the Orcs have a low standing with you and you're wielding a weapon that's got an anti-orc property they can recognize and your armor is made by druids, the enemy of the orcs, then you'll be extremely unpopular with the orcs. If you walk into a town wearing inappropriate clothes or pay with the currency of an enemy faction, that'll be automatically factored into NPC reactions. First impressions would count for a lot, but if you do enough work to get your standings up, your clothes and hairstyle will have less impact in the face of a more fleshed-out "profile" with that faction or individual.
I think the data needed for such interactions could be compressed pretty efficiently, and would never come close to the vast space occupied by texture files in modern games. How much space is occupied by all the relationship data in a Sims game?
@Nomen Nescio;
I don't want to speak for the OP, but as I read it the proposed system is for 'higher' knowledge than directly steering a FPS bot. At least, that's what I was talking about. Rather than deciding to throw a grenade if enemy disappears around a corner, this system might decide wether or not a character likes another or other such high-level concepts. Then another system might step in and act on those decisions. OP might want to correct me on this though...
As for AI being static in nature, that is 'solved' by including elements of randomization etc. Most game AI are as much about the perception of intellligence as the actual thing. Loads of scientists are trying for the real thing, and I have a hard time seeing the first 'true' AI to crop up in the computer game domain.
I don't want to speak for the OP, but as I read it the proposed system is for 'higher' knowledge than directly steering a FPS bot. At least, that's what I was talking about. Rather than deciding to throw a grenade if enemy disappears around a corner, this system might decide wether or not a character likes another or other such high-level concepts. Then another system might step in and act on those decisions. OP might want to correct me on this though...
As for AI being static in nature, that is 'solved' by including elements of randomization etc. Most game AI are as much about the perception of intellligence as the actual thing. Loads of scientists are trying for the real thing, and I have a hard time seeing the first 'true' AI to crop up in the computer game domain.
All the best internet security is developed for and financed by porn sites, and all the best real-time graphics and AI systems show up first in video games. It's how the world works.
Once the tech is developed, though, I think it could translate easily to FPS games. Security cameras and surviving enemies report on your tactics, so if you use a lot of flashbangs, they'll start to expect it, and get better at closing their eyes and covering their ears when they hear the nade clatter into the room. You get a reputation for using powerful weapons with little penetration, and they'll get behind cover and wear body armor rather than move around a lot. You use a lot of gas, they'll start deploying troops with masks. You use stealth, they'll get flashlights, and on and on and on. Smartening up the enemies is way better than just giving them more HP or sending larger waves. Sure, it's easy to take on a sniper if you're invisible and can use bullet-time to dodge his shots, but that same setup will be far less effective against a guy with thermal goggles and an uzi.
Once the tech is developed, though, I think it could translate easily to FPS games. Security cameras and surviving enemies report on your tactics, so if you use a lot of flashbangs, they'll start to expect it, and get better at closing their eyes and covering their ears when they hear the nade clatter into the room. You get a reputation for using powerful weapons with little penetration, and they'll get behind cover and wear body armor rather than move around a lot. You use a lot of gas, they'll start deploying troops with masks. You use stealth, they'll get flashlights, and on and on and on. Smartening up the enemies is way better than just giving them more HP or sending larger waves. Sure, it's easy to take on a sniper if you're invisible and can use bullet-time to dodge his shots, but that same setup will be far less effective against a guy with thermal goggles and an uzi.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement