Question on AI difficulty

[pasternakski]

ORIGINAL: byron13

Yeah, but you really can't expect a company with resources as limited as Matrix has to be the company that designs the first intelligent AI. I'm just wondering if there is a kind of generic intelligent AI that can be developed that could be adapted to almost any game - or any strategic game. That way, one company could develop it and then license it. But, I admit I don't know - unless I'm arguing with mdiehl. [:'(]

Does the word "consortium" ring a bell? All stand to profit from this area of mutual endeavor.

BTW, if you need help setting this up, I know where you can get a lawyer cheap (not to be confused with a cheap lawyer).

[ZOOMIE1980]

ORIGINAL: Joel Billings

We have removed combat cheats from the hard level. At the "historical level", the AI has a few rules it does not have to follow (due to it doing things differently than the human player. At the hard level, it gets various benefits like increased production, better pilots, etc., but it does not get a combat formula cheat. At the very hard level, it gets everything at hard level (sometimes with more benefit) and it gets some combat cheats (but less than in UV). I suggest everyone play at hard level after their first game (very hard once you really know the game well). You will not notice the benefits the AI gets, but it will help the AI out. The AI does not get smarter at any level. These games are so complex that the smartest we can make the AI is still dumb compared to any decent player after their first game.

Well, that's kind of what I was afraid of. The AI just gets more and better stuff at harder level. Too bad, really. Not everything that makes an AI better has to be hard. For instance, instead of having one broad, general strategy with generally fixed reactions to the typical things a player does, one can randomly choose from about a half dozen quite different strategies.

Take the Japanese attack strategy after Pearl Harbor. One AI path follows the historical path. Another forsakes a drive into the Solomons and New Guinea and puts everything into Burma in an all out attempt to take India and force the British complete out of the game. Or China. Or another that attempts to actually take Pearl Harbor by stealing divisions from Manchuria that were just sitting out the war watching the Russians. And so on.

Maybe on the harder level, the Japanese simply get even more aggressive with their carriers and spend much of the last half of 1942 hunting down what's left of American carrier force, sink it, and then attack Midway and actually attempt to take Pearl Harbor using their entire carrier fleet, before Amercian production kicks in. Or hard level strategy II is to take Australia. Maybe the Japanese actually switch tactics at harder level and attack American transports in force. Different sub deployment strategies, etc.... Basically simple stuff. Tedious and time consuming programming, but not particularly technically difficult, requiring Doctorate level theory.

And a random set of shorter term tactics instead of the same general tactic every time.

I'm willing to bet there is not a dedicated AI developer on this project. One who's sole purpose of employment is to create the AI. A game this grand seemingly would have warrented as much effort put into the AI as was put into the historical accuracy and combat resolution engine.

[ZOOMIE1980]

ORIGINAL: byron13

Yes, high AI levels ideally would do nothing more than dial up the competence of the opponent so that it makes better use of its resources; make it more experienced, if you will. The problem is developing relatively cheap AI that either (i) learns by itself or (ii) is provided with "intelligence" by providing an enormous amount of logical decisions. Hasn't been done yet, and Matrix is not going to be the one to do it.

I think giving the AI more intelligence on your movements at a higher setting would also be relatively expensive to code. Of course, I have no idea how the computer "sees" you now or reacts to units seen. But I'm not sure it is as easy as doubling its search range. In any event, just "seeing" that you are massing troops and carriers in the South Pacific as opposed to the Central Pacific may or may not be sufficient to trigger a response - depends on how the AI is already coded. Or, if it "sees" that all of your carriers are damaged and sitting in San Diego, does this trigger a response? I simply don't know how they've limited the AI when it is suffering from the Fog of War. This would determine what would happen when the fog is lifted.

The easiest way to dial up the competition is simply to give the computer more stuff or cheat on combat results, and that's what they do.

Maybe "smarter" is the wrong term. I'm thinking that "smarter" is actually nothing more than a bit less predictable.

I remember back in the Mid 1980's when SSI came out with North Atlantic 86 for the Apple. Great game back then, but after a couple of runs through the game, all I had to do was place all my missle launching subs up north of England, pull out my Exocet equipped surface ships from my British fleet at Iceland, send the rest to America for later use, and then just wait for the inevitable Soviet amphibious assault on Iceland, all the while putting all my best planes (Tomcats, Tornadoes, and Eagles) on Iceland, maximize supplies and three days before their landing start moving in 25,000 new troops. Worked every time because the AI did the same exact thing every time. A simple little additiion of having the Soviets randomly launch an all out, unsupplied Airborne assault in Iceland on turn 1 instead would have added infinetly to the unpredictability of the solitare version of the game. And added only about 400 lines of Applesoft Basic code! hell even two or three random paths by the amphibious assault TF's instead of the one they always used so you couldn;t just mass your subs at one spot.

I just don't want an overly predictable AI. Quite frankly, the AI in UV is no better than the AI in the 1985 version of North Atlantic 86. And that's a sad comentary on the state of AI development in gaming these days.

[mogami]

Hi, The AI would have to be able to form it's own long range plans and then know when and how to modify them in reaction to what the human player did. It has taken years to program an AI to play chess to where it can compete with human grand masters (but only at the fast time controls) Thats a game with 16 pieces and 64 squares. Each move in chess can be calculated out depending on time allowed and speed of processor. (looking 10 moves ahead takes all night)
In war games the vast majority of opposing pieces are out of sight.
In chess one human advantage is knowing which possible future series of moves are worth further study and which can be discarded with only a second or two of thought. The computer must calculate each and every possible tree and assign a numerical value to each and when out of time use the move accorded the highest score. What I'm hinting at here is to get an AI that could compete with a knowledgeable human it would need to spend several days on each turn. (just thinking) AI's tend to do things that don't make sense because they are not following a plan or reasoning things out. They are following instructions programed into them ("Capture Rangoon, defend Truk") These instructions may have nothing to do with the actual events taking place in the game. And humans learn the AI habits and then exploit them and then complain the Ai is too stupid. Once you learn the AI only sends Tf's with 2 CV and maybe 1 or 2 CVL stop fighting it with 6-8 CV. (Try to capture Truk and defend Rangoon)
Don't place a surface combat TF in the hex the AI always uses when moving it's transports. (pretend you don't know what and how the AI is doing ) But really you can't eat your cake and have it too. If you want to play a fast game play the AI on hard settings. If you want a tough challanging game find a human. This is not a real time game. You are going to want to spend a lot of time doing your turns so you have to allow your oopponent to do the same. I think 1 turn of WITP per day is a lot of WITP.
To really spice things up start 1 from each side with the same player. Then you just trade turns and your always busy.

[ZOOMIE1980]

ORIGINAL: pasternakski

ORIGINAL: Damien Thorn
I'd rather give the computer more stuff than get jumped everytime I turn around because the computer knew exactly where I was going.

But, see, that's the whole problem. If there is an AI, it shouldn't need more stuff or better intelligence. It should work better with what it has.

It just needs to be less predictable. Do the Japs always have to move, enforce, into the Solomons? Do the Americans really need to always counterattack so forcefully in Solomons, vs putting more effort into getting to the Mariana's fast and cutting Rabaul and Solomons complete off, earlier?

Obviously on Historical settings you want historical strategy, but even within the Historical settings the AI can be less than totally predictable. What's wrong with the AI making a surprise Carrier attack on Brisbane, for instance, or have the Japanese feint into Darwin or Cairnes to maybe lull the Americans to mistakenly commit a "loaner" division to Austrailia and screw with their timeline on Guadalcanal? Or the same in the Burma theator. Maybe on the hard or very hard setting the Japanese fient into India with a surpise landing of sacraficial division into Dacca or even Calcutta to get the British and Indian forces to panic and pull a division or two out of Burma. And if they they don't panic, turn it into a REAL end run assault to totally flank the whole theater?

That's the kind of stuff I expect out of crafty AI.

[mogami]

Hi, Always give the AI the easiest side. (the defending side normally)

[ZOOMIE1980]

ORIGINAL: Mogami

Hi, The AI would have to be able to form it's own long range plans and then know when and how to modify them in reaction to what the human player did. It has taken years to program an AI to play chess to where it can compete with human grand masters (but only at the fast time controls) Thats a game with 16 pieces and 64 squares. Each move in chess can be calculated out depending on time allowed and speed of processor. (looking 10 moves ahead takes all night)
In war games the vast majority of opposing pieces are out of sight.
In chess one human advantage is knowing which possible future series of moves are worth further study and which can be discarded with only a second or two of thought. The computer must calculate each and every possible tree and assign a numerical value to each and when out of time use the move accorded the highest score. What I'm hinting at here is to get an AI that could compete with a knowledgeable human it would need to spend several days on each turn. (just thinking) AI's tend to do things that don't make sense because they are not following a plan or reasoning things out. They are following instructions programed into them ("Capture Rangoon, defend Truk") These instructions may have nothing to do with the actual events taking place in the game. And humans learn the AI habits and then exploit them and then complain the Ai is too stupid. Once you learn the AI only sends Tf's with 2 CV and maybe 1 or 2 CVL stop fighting it with 6-8 CV. (Try to capture Truk and defend Rangoon)
Don't place a surface combat TF in the hex the AI always uses when moving it's transports. (pretend you don't know what and how the AI is doing ) But really you can't eat your cake and have it too. If you want to play a fast game play the AI on hard settings. If you want a tough challanging game find a human. This is not a real time game. You are going to want to spend a lot of time doing your turns so you have to allow your oopponent to do the same. I think 1 turn of WITP per day is a lot of WITP.
To really spice things up start 1 from each side with the same player. Then you just trade turns and your always busy.

I think you're making it harder than it really has to be. I'm mostly talking about some fairly basic, simple, stuff to make things a little less predictable at higher AI difficulty levels. For instance, take the AI playing the Japanese.

There's not a great deal the human player can do to the Japanese before late summer 1942. The American and British just don't have much to work with until production sets in. But by May 1942, the Japanese have their pre-war planned defensive line set up. They cut the Burma road, have most Eastern, coastal Chinese cities under control, Singapore, The East Indies with their oil, Rabaul and the North coast of New Guinea, the Marianas and Wake and Tarawa. All that is fine. The historical model moves from there, historically. The Japanese choice from there is 4 carrier thrust at Midway with a following assault force for the base, and thrust south in the Solomon's and a feeble attempt at rounding the east end of New Guinea to get to the south coast.

That's where a "smarter" AI could take over on the hard or very hard level. Instead of the more and better, how about an alternate strategy. Have the AI add two more carriers and another division to the Midway assault and forget the southward thrust in the Solomons. Or forget Midway and put everything into taking Port Morseby. Or have the AI grab many of those Manchurian Divisions and make an end run assault on Dacca ala McArthur at Inchon a few years later. Or use them to finish off China once and for all. And have a relatively random chance at which tact it takes so the human player is caught off guard sometimes.

Or something as simple as massive eight carrier 2nd Pearl Harbor attack in Aug of 1942 or something shocking like that (at the hard or very hard level). Or have the Japs make a surprise attack on American Somoa and really screw with the Australian supply line. And make it random, so you're not sure if they are ever really going to do it.

Stuff like that.

[brisd]

I understand where you are coming from. By AI, you are referring to the pre-programing instructions of the computer player, not an vastly improved computer player able to 'think and learn'. I believe in classic Pacific War the computer player would have random grand strategies pre-programed in. I think both UV and WITP too will have what you want. It is just after a few plays, the average and above player will figure out what's up. I too like to play solitare and want a decent computer player. Against a real live opponent via pbem is the true challenge. And I remember playing North Atlantic 86 too! Ah, we are so spoiled these days with our pcs and games like WITP. [8D]

[ZOOMIE1980]

Yea, just a degree of randomness, so you can't just sit in a predicted hex and wait, or massively fortify a particular position knowing full well the AI is going to attack right there, each and every time.

[Capt. Harlock]

I remember back in the Mid 1980's when SSI came out with North Atlantic 86 for the Apple. Great game back then, but after a couple of runs through the game, all I had to do was place all my missle launching subs up north of England, pull out my Exocet equipped surface ships from my British fleet at Iceland, send the rest to America for later use, and then just wait for the inevitable Soviet amphibious assault on Iceland, all the while putting all my best planes (Tomcats, Tornadoes, and Eagles) on Iceland, maximize supplies and three days before their landing start moving in 25,000 new troops. Worked every time because the AI did the same exact thing every time. A simple little additiion of having the Soviets randomly launch an all out, unsupplied Airborne assault in Iceland on turn 1 instead would have added infinetly to the unpredictability of the solitare version of the game. And added only about 400 lines of Applesoft Basic code!

I remember the game as well: in fact I managed to print out the "source code". (It was uncomplied Applesoft BASIC.) 400 lines was not possible; remember we're talking about a program that must fit into 48K of RAM, and come from a 143K floppy.

But now we have huge amounts of RAM and hard disk space available, so we should be able to have a genius-level AI, right? Well, no, we still have major limits in the ability of the programmers to produce a game in a reasonable time and at a reasonable cost.

However, let me elaborate a little on my idea of giving the AI better intelligence. There comes a point in the game when the Japanese side needs to switch to playing defense, and the Allied side should switch to offense. Perhaps the higher level AI could be given a more accurate calcualtion for when that time is.

[UndercoverNotChickenSalad]

I think there is a certain incompetance humans have that is impossible to program [:'(] Something about pros being predictable but a newbie might do something so ridiculous that it catches you totally by surprise [8|]

[mogami]

Hi, One of the great myths with non chess players is that by doing the unexpected you can confuse the masters. There are reasons for a person becoming a master. First they understand the game and know what is possible and what is just silly. Making crazy moves against a master is a certain means of losing badly. Against a master you need to make simple logical solid moves. You need to know why you are making your moves and what the position will and will not allow. You can't "outsmart" or "confuse" the master all you can do is not allow him any weakness or crack in the position he can exploit. The master understands the only valid reason for an attack is a weakness in the enemy position that he can operate against.
It is not possible to foresee what positions might arise 400 turns into a 1600 turn game. The AI cannot be programed in advance to do "unexpected" things. "Unexpected" usally means unjustified, risky and plain old stupid.
To make the AI "smarter" it would need to be programed to assign a numerical value to all the possible situations in existence at the start of the turn (it would need to recount each and every turn) It would then need to examine all the possible changes (each and every turn)
Making the AI able to this is not hard. But then you would only be making 1 turn every 3 days and the computer would need to be allowed to stay on so the AI could think between turns.
Preprogramming strategy will only produce a set of human players that have learned to exploit the plans. The plans would have no bearing on the actual course of the war.
A few examples: Some Allied players will run from the SRA when the war begins. They will transfer all the airgroups to bases the Japanese will not overrun. All the ships will run to Ceylon or Darwin. LCU's will be evacuated where possible. This Allied player will remain only at a few bases far from Japanese power and wait for reinforcements to arrive. Risk nothing and go slow.
Another Allied player will mass what forces are present in SRA at start and look for a target he can defeat. He will hope to catch underescorted Japanese TF's outside Japanese LBA but he will fight.
A third Allied player will direct everything in SRA to attack the nearest Japanese target no matter what.
The non smart AI will need to have instructions to handle each type and it will need to know how to recognize what type it is encountering and how to tell if the opponent changes style. (Style is a good word. I think instead of historical hard and very hard the settings should be called select AI style because the historical level only refers to weapon accuracy and effect and not the operational course of the war. Styles would be Cautious, Balanced and Aggressive. Cautious means the AI will not be likely to take risks and will strive to have a numerical advantage at points of contact. Balanced will be a mix of the two and aggressive means the AI will only be interested in killing the enemy no matter what the cost (it will still look for the safest means of doing this but it will direct it efforts towards attacking the human player where it feels he is the weakest)
Here no pre instructions are required but the human player will have to select how much time the AI gets per turn to think. The more time allowed the better the AI should preform. (all other things being equal the more time you give a calculator to calculate the better it's calculations will be)

[ZOOMIE1980]

Once again you make a poor case. I've programmed various AI algorithms for over 20 years. You have the basis mostly correct, but the resourse claims are ludicrous. The notion that it would take an AI 3 days per turn to assign value to all the possible actions it would need to take each turn is ridiculous. The notion that a "smart" AI need consider 400 turns in advance is also nonsense, even against a "master". Also completely undefined is just what, in context of this game, is a "master". How long does it take a player to play a 1600 turn game to become a "master"? At best, a few players will play enough to get "good". Most will get to be merely reasonably competent.

I absolutely guarantee, one dedicated professional could develop a chess-like AI that could look 15-20 or so turns in advance and give 80-90% of all the people that will ever buy the game, a truely challenging game against a very intelligent API and be able to do with the AI taking more than several minutes on a PIV 3.2GHZ machine, to calculate it's best response. Certainly less than hour or so. Much less for lower intelligence settings.

But even forgoing that genre of AI, while tedious, a dedicated developer could develop enough preprogrammed strategies each with several possible strategic branches during the course of the game, based on general human opponent strategy, to keep even proficient players somewhat off gaurd, and keep the AI from being overly predictable.

What I continue to see from the development team here is an attempt to justify why the AI is well down the priority list of the game. Hopefully, after the realse, if successful enough with enough cash flow, that someone will decide to put some significant effort into an a major AI upgrade patch down the road.

[pasternakski]

And please take a look at what Brad Wardell's people have done in terms of having the AI working in the background while you take your turn. I think that this may be one of those revolutionary breakthroughs we are looking for.

[mogami]

ORIGINAL: ZOOMIE1980

Once again you make a poor case. I've programmed various AI algorithms for over 20 years. You have the basis mostly correct, but the resourse claims are ludicrous. The notion that it would take an AI 3 days per turn to assign value to all the possible actions it would need to take each turn is ridiculous. The notion that a "smart" AI need consider 400 turns in advance is also nonsense, even against a "master". Also completely undefined is just what, in context of this game, is a "master". How long does it take a player to play a 1600 turn game to become a "master"? At best, a few players will play enough to get "good". Most will get to be merely reasonably competent.

I absolutely guarantee, one dedicated professional could develop a chess-like AI that could look 15-20 or so turns in advance and give 80-90% of all the people that will ever buy the game, a truely challenging game against a very intelligent API and be able to do with the AI taking more than several minutes on a PIV 3.2GHZ machine, to calculate it's best response. Certainly less than hour or so. Much less for lower intelligence settings.

But even forgoing that genre of AI, while tedious, a dedicated developer could develop enough preprogrammed strategies each with several possible strategic branches during the course of the game, based on general human opponent strategy, to keep even proficient players somewhat off gaurd, and keep the AI from being overly predictable.

What I continue to see from the development team here is an attempt to justify why the AI is well down the priority list of the game. Hopefully, after the realse, if successful enough with enough cash flow, that someone will decide to put some significant effort into an a major AI upgrade patch down the road.

Hi, I don't think I suggested that the AI would need to look 400 turns in advance. I think I said a programmer could not pre programe strategy 400 turns in advance because he will not know what transpires in those 400 turns.

Why don't you help out and post strategy they can pre program for the AI to play Allies or Japanese for scenario 15 (the complete war Dec 41 to June 46)
It would help if you could define all possible courses the war could take because having the AI following a pre planned stategy that has no relation to the current on map situation is worse then having a merely stupid AI.

Also I was not saying there were or would ever be masters of WITP. I said novice chess players that think they can confuse master chess players by making "unexpected" moves lose the game faster then if they stick to simple but solid moves. It does not matter if the master can predict them as long as they do not present him with a weakness that he can then exploit to his advantage. A master will see and recognize a weakness and "unexpected" moves are where they generally occur.

I don't mind being rebutted but I do mind be rebutted using points I did not make. I'm sorry if I was unclear.
I play high dollar ,high rated chess machines. At their fastest settings they are all at least 400 points (USCF rating) below what they play at their slowest settings. Just don't play where you try to out calculate them. Play solid closed long term positional chess and watch them suffer breakdowns. (Rookies open the position and then try tactics against a computer) Also most chess programs have an opening libary where the past 300 years of grandmaster chess is stored. The program just checks for positons that match one from it's libary. (so it is not really understanding or thinking just plagerizing history) It is illegal for a human player to consult such a libary during a game. In matches where the libary is removed the program has a dramamatic drop in results. Chess programs have scored their highest in "Blitz" games where they can use their libaries and calculating powers in tactical games to advantage. They remain almost hopeless when deprived of the libary and are forced to play slow closed strategic games. (They are suckers for material sacrifice and will never be able to understand the positional nature of such sacrifices because the grandmaster can not explain it only knows from experiance what it means and how to exploit it. )

[Mike Scholl]

MOGAMI One thing you are forgetting in your arguments here is that compared to
WITP, Chess is a VERY SIMPLE game. I'm not talking in terms of strategy, just in
terms of possibilities. Chess you move one piece at a time, which can interact (take)
at most one piece at a time. And can move in a few well-defined and limited ways.
In WITP, every asset on a on both sides can be "in play" at the same time, some con-
strained by various forms of geography, some not. Some pieces "carry" other pieces.
The combine in different ways in attacks on one or several other pieces at the same
time.
Good Chess programs are available, Great ones are possible but not readily available.
But a just "good" AI Program for WITP would need about 100 times the resources as
a "great" chess program just to keep up with the millions of possible variations that
can arise. It's a neat goal, but not realistic currently.

[pasternakski]

Well, let's think about the complexity of chess for a moment. There are only 32 pieces, it is true, and only 64 squares on which they can stand.

What makes chess a tremendously complex game for the computer to play is the range of possibilities over the course of several moves. Human players discard tens of millions of those threads of play without even thinking about it. The computer, on the other hand, has to consider every single one of them.

WitP is a very different proposition, it seems to me. Each "piece" has a limited role and capability. The computer is not really looking at a game map and is not calculating on the basis of what it "sees" and "knows." It is merely applying coded formulae in lockstep fashion to generate a result that can be translated, through the game's graphics, into something the human player can see and interpret. This makes the AI very much a closed system. What I suspect can be done to improve AI play is to loosen the formulae that are applied and add alternative sequences that can be triggered either by the existence of certain preconditions or, in some instances, even randomly.

Of course, I don't have the slightest idea of what I'm talking about.

[Mr.Frag]

Of course, I don't have the slightest idea of what I'm talking about.

[:D]

Worthy of a tag line!

[Mike Scholl]

ORIGINAL: pasternakski

Well, let's think about the complexity of chess for a moment. There are only 32 pieces, it is true, and only 64 squares on which they can stand.

What makes chess a tremendously complex game for the computer to play is the range of possibilities over the course of several moves. Human players discard tens of millions of those threads of play without even thinking about it. The computer, on the other hand, has to consider every single one of them.

Of course, I don't have the slightest idea of what I'm talking about.

And if WITP had only 64 hexes, and 16 pieces on a side, it WOULD STILL BE MUCH MORE
COMPLICATED for the computer. Instead of looking for the "best move" for one piece,
it has to find the best move for ALL of it's pieces, some of which will be acticing in unison
to accomplish a single mission. Then it has to analyze the other side's potential moves
for ALL of it's pieces, and how the potentialities will relate to what it's trying to do. As
WITP has a LOT more "squares", and a LOT more "pieces" Not to mention that ALL of
your opponants peices can move simultaneously with ALL of yours! So while you were
probably making a "funny" I think you got your last statement correct.

[Mr.Frag]

Don't forget the obvious point ... there is no fog of war in chess.

You get to see all the information all the time and there are only 6 pieces with very specific rules of movement, not 16 on each side ... a rook is a rook is a rook. Does not matter who owns it as it does not move differently for white or black.