The truth set us free.trepan wrote:Or the truth?
AI improver gadget - new version 1.01
Moderator: Moderators
DJ, while it has been said by others, I feel I should stress it. The capability of the human mind for creativity is the greatest source of skill in an open environment, and it is also the most difficult thing to replicate in artificial intelligence because our understanding of logic is intrinsically vague and thus in many ways cannot be represented mathematically. Any AI routine, however, must act in a rational, mathematical framework. Thus, you compensate for the impossibility of modeling creative thought by using simple mathematical cheats, increasing speeds of simple calculations, etc.
OK i accept what you are saying but there are areas in which an AI has an advantage, particularly in RTS games. These are what we should be looking to exploit in AI design.
Micro management is perhaps the best area, an AI should be able to use teams of artillery or other long range weapons to their greatest effect. Where as a human can only micro manage a small number of teams effectively an AI is only limited by the capacity of the CPU it is run by.
KAI is perhaps the best AI we have at present, yes i know it cheats by means of the cheat interface but if I'm not mistaken, it doesn't increase its resources it merely uses the cheat interface to locate enemy units. This could be achieved with an effective scouting algorithm, at least to some degree. The reason KAI is effective is because it micro manages well and identifies objectives -undefended portions of the human players base which it then attacks. The other AI's tend to concentrate on a specific point. Making them easy to defend against. Whereas KAI attempts to stop a human player from expanding.
I haven't said, nor do I believe that AI's can realistically beat human players (look at the power of computer needed to beat a human at chess) but I do think that we can improve our AI's based on work that has already been done within this project. Imagine an AI with KAI's targetting and micro management, combined with the configuration available to NTai and the base building of RAI. It would be a big step forward and all those components have already been developed. Yes it's not as simple as just bolting them together but the principles could be combined into a very strong AI.
I don't want to criticise the efforts of the AI developers at present, I could not have achieved what they have but it seems to me that there are no new ideas going into our AI's at present and nobody pushing the boundries of what has been done. This is because all the AI's are side projects, and as a consequence development is mostly keeping them running. Or in some cases trying to get them running again. Ideally we need someone to make an AI their main target. This is what my post was for.
Micro management is perhaps the best area, an AI should be able to use teams of artillery or other long range weapons to their greatest effect. Where as a human can only micro manage a small number of teams effectively an AI is only limited by the capacity of the CPU it is run by.
KAI is perhaps the best AI we have at present, yes i know it cheats by means of the cheat interface but if I'm not mistaken, it doesn't increase its resources it merely uses the cheat interface to locate enemy units. This could be achieved with an effective scouting algorithm, at least to some degree. The reason KAI is effective is because it micro manages well and identifies objectives -undefended portions of the human players base which it then attacks. The other AI's tend to concentrate on a specific point. Making them easy to defend against. Whereas KAI attempts to stop a human player from expanding.
I haven't said, nor do I believe that AI's can realistically beat human players (look at the power of computer needed to beat a human at chess) but I do think that we can improve our AI's based on work that has already been done within this project. Imagine an AI with KAI's targetting and micro management, combined with the configuration available to NTai and the base building of RAI. It would be a big step forward and all those components have already been developed. Yes it's not as simple as just bolting them together but the principles could be combined into a very strong AI.
I don't want to criticise the efforts of the AI developers at present, I could not have achieved what they have but it seems to me that there are no new ideas going into our AI's at present and nobody pushing the boundries of what has been done. This is because all the AI's are side projects, and as a consequence development is mostly keeping them running. Or in some cases trying to get them running again. Ideally we need someone to make an AI their main target. This is what my post was for.
No, it isn't logic we don't understand (that has
been studied for 2000+ years, and is pretty well
formalised). The problem is capturing creativity
and uncertainty within a formal logic reasoning
system in a consistent way, and that's where you'll
find opinions differ on what is and isn't possible
(one of the reasons why the field of AI exists
).
ps. as far as this gadget goes, it's simply a more
advanced form of issuing handicaps, so I don't really
see much wrong with it.
been studied for 2000+ years, and is pretty well
formalised). The problem is capturing creativity
and uncertainty within a formal logic reasoning
system in a consistent way, and that's where you'll
find opinions differ on what is and isn't possible
(one of the reasons why the field of AI exists
).ps. as far as this gadget goes, it's simply a more
advanced form of issuing handicaps, so I don't really
see much wrong with it.
-
Forboding Angel
- Evolution RTS Developer
- Posts: 14673
- Joined: 17 Nov 2005, 02:43
THis thing is awesome. Rai handed my ass to me on a silver platter. I even nuked it (the nukes in funta are really hard to get and have an area affect of 2500), I didn't even slow it down 
I was using my doomsday laser like they were going out of style, it just smiled and kept on coming. God, Immolator towers, for all their short range, are devastating when they catch you by surprise. My poor amphibs didn't even have a chance. Remind me to add stealth and cloaking to amphibs so torps can't get at them so easily.
I set the rush protection to 90 seconds (so that any AI would have a very strong start). The battle raged for about 2 hours I think, but finally, the RAI broke through my air defenses and cracked me like an Egg.
Man that was 100% made of win. Thank you bighead, I have wished for something like this since forever.
I was using my doomsday laser like they were going out of style, it just smiled and kept on coming. God, Immolator towers, for all their short range, are devastating when they catch you by surprise. My poor amphibs didn't even have a chance. Remind me to add stealth and cloaking to amphibs so torps can't get at them so easily.
I set the rush protection to 90 seconds (so that any AI would have a very strong start). The battle raged for about 2 hours I think, but finally, the RAI broke through my air defenses and cracked me like an Egg.
Man that was 100% made of win. Thank you bighead, I have wished for something like this since forever.
-
Forboding Angel
- Evolution RTS Developer
- Posts: 14673
- Joined: 17 Nov 2005, 02:43
Actually he's right.
KAI NTai and QAI dont use a maphack because its necessary, they use it to save time and effort, basically we AI developers wanted to save time so we didnt bother fixing/writting scouting algorithms and instead took a shortcut and just used a maphack.
Efficient scouting algorithms do exist and have been sued in the past. JCAI did quite well with scouter aircraft, and my own mex spot based scouter algorithm did very well with raider type units early on.
KAI NTai and QAI dont use a maphack because its necessary, they use it to save time and effort, basically we AI developers wanted to save time so we didnt bother fixing/writting scouting algorithms and instead took a shortcut and just used a maphack.
Efficient scouting algorithms do exist and have been sued in the past. JCAI did quite well with scouter aircraft, and my own mex spot based scouter algorithm did very well with raider type units early on.
-
Forboding Angel
- Evolution RTS Developer
- Posts: 14673
- Joined: 17 Nov 2005, 02:43
-
KingRaptor
- Zero-K Developer
- Posts: 838
- Joined: 14 Mar 2007, 03:44
On large maps it seems KAI fails terribly since it just builds army and doesn't attack. Comet Catcher was a waste, though that 90% complete Vulcan totally caught me by suprise... 
In other words, we need an AI that actually is designed to use this to its advantage. That would be nothing less than sick.
In other words, we need an AI that actually is designed to use this to its advantage. That would be nothing less than sick.
-
Warlord Zsinj
- Imperial Winter Developer
- Posts: 3742
- Joined: 24 Aug 2004, 08:59
I actually don't mind cheating AI's at all; as long as the cheating is very subtle; and you're always in two minds as to whether the AI is really cheating or is just really damn clever/lucky. I always hate seeing cheating AI's that are quite clearly cheating, doing impossible things, it really ruins immersion.
The best cheating AI that I ever played was in Shogun: Total War.
In that game, the AI was already very good, but it also cheated on the hardest levels. But I only conclusively knew it cheated much later, when I was told about it; I couldn't actually prove it was cheating while I was playing.
How it cheated, apart from getting more resources, was that, they always had a few more arrows then you, so that if you got caught in a shooting war, they would usually come out on top, because they could fire for longer. They also had more stamina, so that if the battle went for longer, they would win.
This meant that you had to go out of your way to defeat the enemy, otherwise you would lose. It made the game fascinating, as it put the onus on you to be clever and defeat the AI; and it forced you from doing what most people like to do with AI's, which is sitting around with a huge fortress and waiting as the AI exhausts themselves on your walls.
So yeah, cheating AI's are fine, in my opinion, as long as you handle the cheating with finesse and thought, rather then just making them insanely powerful.
The best cheating AI that I ever played was in Shogun: Total War.
In that game, the AI was already very good, but it also cheated on the hardest levels. But I only conclusively knew it cheated much later, when I was told about it; I couldn't actually prove it was cheating while I was playing.
How it cheated, apart from getting more resources, was that, they always had a few more arrows then you, so that if you got caught in a shooting war, they would usually come out on top, because they could fire for longer. They also had more stamina, so that if the battle went for longer, they would win.
This meant that you had to go out of your way to defeat the enemy, otherwise you would lose. It made the game fascinating, as it put the onus on you to be clever and defeat the AI; and it forced you from doing what most people like to do with AI's, which is sitting around with a huge fortress and waiting as the AI exhausts themselves on your walls.
So yeah, cheating AI's are fine, in my opinion, as long as you handle the cheating with finesse and thought, rather then just making them insanely powerful.
The saddest things about cheating AI's is the 'anti rush' and starting with an impossible amount of resources. Against a real player, in most mods, attacking early and hard is vital to stop your enemies expansion- and yes, if you do it will enough, you'll win the game then and there.
Yet AI's tend to favour a war of attrition approach towards them, so that the way to win is porcing up and waiting for them to crack (Though i cant really speak for Spring AI's- i dont play against AI's).
Thats a shame, as it really teaches new players terrible habits.
Yet AI's tend to favour a war of attrition approach towards them, so that the way to win is porcing up and waiting for them to crack (Though i cant really speak for Spring AI's- i dont play against AI's).
Thats a shame, as it really teaches new players terrible habits.
New version is out!
AIs are now aware of increased income. This makes them fight much better in long games (and big maps). AIs also cheat less. Especially at the beginning - AIs start with 2x more resources instead of 4x.
It's now possible to survive t1 spam (well, you still need to micro like mad) and have long and interesting game. I've just had one
AIs are now aware of increased income. This makes them fight much better in long games (and big maps). AIs also cheat less. Especially at the beginning - AIs start with 2x more resources instead of 4x.
It's now possible to survive t1 spam (well, you still need to micro like mad) and have long and interesting game. I've just had one
In this version of Spring, players would be forced to use it. Once new version of Spring is out I will be able to release accompaning "difficulty settings" widget.
Plans for next version:
Nice lua dialog for hosting player (yes, it will pop out only if there are AI teams):
The game is paused. Please choose your difficulty setting:
Easy (no cheating)
Medium
Hard
Last edited by BigHead on 11 Nov 2007, 18:04, edited 1 time in total.
-
Pressure Line
- Posts: 2283
- Joined: 21 May 2007, 02:09