Game development and why most games have "bad" AI

[Kriegsspieler]

Well, for what it's worth, I have always been struck by the fact that P'dox positions its games in a niche we could call "historical sandbox", which I swear must be the most difficult of all to produce a workable a.i. for! How do you get an a.i. that both behaves in plausibly "real" ways for the historical scenario envisioned in the game yet does not follow a tightly scripted sequence of actions? All of their games require this balance in one way or another.

Another discussion of a.i. got me to thinking about why the a.i. in War in the Pacific Admiral's Edition is so good (and it is good, if you've never played that monster ). I think the reason is that the historical scenario is so tightly controlled. The war is between two parties, the Allies and the Japanese. You can actually "turn off" China, if you want to -- and although most of the units in China have to stay there, in order to avoid upsetting the strategic situation in the rest of the Pacific war, those units sit there and don't do much of anything. In general, the a.i. acts to establish a defensible perimeter for the Japanese, or collapse it for the Allies. In other words, it does what the historical actors did in the war. It is quite capable of surprising a human player, but does so only within a prescribed set of limits.

I do think that the kind of a.i. thinking represented in HOI3 is a real advance from what we had seen before, although at present its fruits are mostly to be harvested in the future. And here's hoping they're transferable to this game!

 

[Yaromir]

Special AI perks can be infuriating if it allows the AI to do things player, in the same condition, cannot do.

Its understandable AI getting +10% production, or other some such bonus because it can't be taught to manage its resources as efficiently as a human would.

What's not OK, is when an entire game-mechanic doesn't apply (or applies very differently), for example AI not suffering Naval Attrition so it can sail its fleets half the world to blockade your ports.

What I found in CIV IV, was that on higher difficulty levels, AI simply gets a bigger head-start (starts with more units). If you survive to catch up, the game plays roughly same as lower difficulty levels.

 

[Chaingun]

^

Interesting observations! I agree ideally AI *should* behave so that no game mechanic is rendered completely useless (but this ideal is constrained in practice like earlier discussed in this thread).

Also, like you've noted, it's probably better to give the AI a bonus that is a derivative to time rather than a fixed bonus at start. Though theoretically in Civ more units at startgive exponentially scaling advantages, in practice AI is worse at utilizing them than the human player and thus will have a lower exponential rate, which means any constant factors at start are canceled out given enough time. E.g. (10 * 2 ^ 6t) will be outgrown by (2 ^ 7t).

 

[sbr]

Well, for what it's worth, I have always been struck by the fact that P'dox positions its games in a niche we could call "historical sandbox", which I swear must be the most difficult of all to produce a workable a.i. for! How do you get an a.i. that both behaves in plausibly "real" ways for the historical scenario envisioned in the game yet does not follow a tightly scripted sequence of actions? All of their games require this balance in one way or another.

Another discussion of a.i. got me to thinking about why the a.i. in War in the Pacific Admiral's Edition is so good (and it is good, if you've never played that monster ). I think the reason is that the historical scenario is so tightly controlled. The war is between two parties, the Allies and the Japanese. You can actually "turn off" China, if you want to -- and although most of the units in China have to stay there, in order to avoid upsetting the strategic situation in the rest of the Pacific war, those units sit there and don't do much of anything. In general, the a.i. acts to establish a defensible perimeter for the Japanese, or collapse it for the Allies. In other words, it does what the historical actors did in the war. It is quite capable of surprising a human player, but does so only within a prescribed set of limits.

I do think that the kind of a.i. thinking represented in HOI3 is a real advance from what we had seen before, although at present its fruits are mostly to be harvested in the future. And here's hoping they're transferable to this game!

Besides the limited scope of the game, the other advantage the AI has in WitP is that it is turn based; that really helps the AI keep track of what is happening.

 

[Colon]

What I found in CIV IV, was that on higher difficulty levels, AI simply gets a bigger head-start (starts with more units). If you survive to catch up, the game plays roughly same as lower difficulty levels.

I know I'm straying too far from Vic2 right now, but the Civ4 AI does have durable bonuses on higher difficulty levels and those increase with each level, as you can see here: http://forums.civfanatics.com/showthread.php?t=158130

 

[Kriegsspieler]

Besides the limited scope of the game, the other advantage the AI has in WitP is that it is turn based; that really helps the AI keep track of what is happening.

I've often thought the same thing, to be sure, but I'm not convinced that's right. If the a.i. in a P'dox game only "samples" the situation once per day and then responds, isn't that essentially similar to a turn-based mechanism? Isn't that what happens when the game pauses for its daily and monthly "rollovers"?

 

[Chaingun]

I've often thought the same thing, to be sure, but I'm not convinced that's right. If the a.i. in a P'dox game only "samples" the situation once per day and then responds, isn't that essentially similar to a turn-based mechanism? Isn't that what happens when the game pauses for its daily and monthly "rollovers"?

Indeed it is. It's the number of AI frames that vary. (For this discussion, a frame is defined as the smallest unit of time where player input or other events may effect the game world, to which the AI agent must potentially respond. This can be different from graphics frames. ) In a classic turn based game the number of frames is equal to the number of turns (classic as in non-simultaneous TBS). In a real time strategy game player input can occur 60 times per second or more. Paradox games fall in between with a kind of semi-turn system where frames arrive at a user defined rate.

You can "downsample" AI frames, which is what RTS games in general need to do to some extent for many calculations. But then you trade AI quality for execution time.

 

[unmerged(63310)]

Every action an AI makes is somewhere programmed. That is the defining limit of its capability. As already pointed out that means for games with a limited amount of possible actions(chess has 10^50 possible positions but per turn the number is anywhere from 20 at opening of a game to 1-256 or so possible legal moves from the current positions depending on pieces remaining on the board).

Computers are still dumb logic... IE- any parameter it uses in making a calculation is limited to the programmers ability to know what is important information and what is not since the computer looks at all information as equal. This means the programmer must assign a value to each bit of information the computer calculates with. That works ok in rigid situations with specific parameters but the more "complex" the situation the more a fixed value can be a wrong value at least for calculating an optimal outcome.

Humans can very quickly eliminated +99.999% of the information in a glance and so while a computer in chess might look at over millions of possible moves in the next 20 turns a human will only examine the best 5 or 6 moves in the next 4 or 5 turns which is beyond the max most humans can simultaneously hold in their heads as it is 100's of possible combos even that limited but grandmasters with lots of practice and pattern recognition of similar positions in past games can roughly calculate.

Where Kasparov for example managed to beat computers up until 1997 was in knowing that the value of pieces changed relative to position on the board and could make a sacrifice that would put the computer in a poor position. Also knowing the computer was programmed to calculate based on optimal moves sometimes a suboptimal move could manipulate the computer into making a mistake. However once processing power allowed the chess AI to look enough moves in advance such strategies no longer worked even with fixed values for the pieces as all the computer did was add the values of the pieces left on the board for every combination of moves 20 turns away and made the move that left the highest value of pieces. In fact the math work done since is beginning to indicate that a perfect game of chess from both sides of the board will always lead to a draw.

Now the amount of possible outcomes for a even a simple computer game with 10 different sides with each side having 10 units and the board is over 10,000 tiles is beyond calculation so the programmers have to decide what information is important for the computer to use and then weight that information. Since there is not always a "best move" in a game that is not rock, paper, scissors the AI has some leeway to still give a challenge but most humans will quickly learn all the "better" moves for most common situations whereas the AI is not learning, just repeating its programmed calculation of weighted values for each instance.

WitP is already a smaller game "board" than HoI3 or Vicky 2 so there is less calculations already inherent in the game design and that saves processing power as calculations for units tactical moves is much reduced. The most difficult part of programming an AI for Paradox games would be where to assign the focus- or so I'd guess since I am not involved nor a programmer- by focus I mean dividing up the processing power available in the most efficient way. So I think this often leads to AI that can tactically move its units well enough as this is the most obvious part of the game. Where it is a struggle is on the more strategic scale of unit production etc... since the AI does a frame snapshot of the current situation when designing production. Build scripts and such devices are a stopgap to get the AI to build in a certain way for a likely future outcome but the major flaw is that a human player does not often play consistently or optimally so any build script is not going to be much better solution than weighted value building and many times worse outcome. IE- AI follows script and build 20% armor, 50% infantry, 20% air, 10% IC/infra... obviously there is problems going to happen almost immediately when war starts and that ratio changes or human 2x builds counters to one of the AI's script based unit builds. If the AI is coded to react to human player changing their own builds- how often does the AI evaluate? Once a month, whenever space opens in que for new production, etc... very difficult to anticipate the possible ways to code AI correctly and even if that part is done perfectly- can the computer spend that much processing power on that while also handling graphics, moving units tactically, handling spying, trade, etc for HoI3 example. For Vicky 2 there are less unit types but much more of everything else and its going to be even more difficult to decide what types of AI decisions deserve the most processing power resources as inevitably the human players will key in on where the AI is weakest and exploit that point not because it makes sense for an overall strategy but because concentrating on that weak point makes the AI more vulnerable everywhere else.

Just look how differently player play multi than vs AI. Most strategy multi games player veterans will try and cover every base a least a bit (unless twitch dominated play is most important which for most RTS these days is the case so micro skills take precedence over tactical planning ability) because any opening can be exploited and leveraged into greater weakness but that leads to resources spread out and outcomes more liquid as players judgment and strategic planning has to account for many more possibilities. Where vs an AI opponent the human player will quickly learn where less resources can be spent relatively safely because the AI is just following its code and to change the weak areas the code would have to change.

So I guess maybe at some point if coding ever is able to be more automated(since right now the time for a human to sit and code is probably the most expensive part of making a game- the labor of programmers is expensive because it is specialized and production efficiency is relatively fixed) then we will see a game that has multiple code sets and can alternate on the fly so a human might be able to exploit an AI weakness momentarily but devising an entire strategy around that weakness would be foolish if the AI could switch to follow different instructions at least somewhat logically in response. This would require quite a bit more processing power and probably a much different game design process than currently.

It is not so much that game developers set out to make a bad AI but making a game that a wide range of computer can run and having a fixed time line and budget to work towards certain decisions have to be made on how to prioritize resources available. Since AI takes the most work for the least tangible results per cost it is natural people will complain since even with 10x the budget that part will always be the weakest part of the game. Not to mention as Sid Meier was saying that often gamers are confused as to what they really want... sure hard core gamers want a challenge but I think what most really want is to feel challenged. Just because a person has played a game for enough hours and figured out how to win does not mean that person wants to suddenly start losing and thus decreasing the value of everything learned so far at a higher difficulty setting. Basically they just want to feel as if everything they learned so far is required to win instead of having the feeling it is possible to win with half of what they know. They still want to win just like the noobs but for a game to be successful the noobs going to the the majority of people buying it which means the core difficultly setting will be catered to that group. Anything extra is likely to be simply cheats or bonus awarded to the AI- not entirely new coding instructions.

Hard core gamers probably need to make house rules and have the discipline not to break them if they want more of a challenge. I know that is difficult though... since skilled players want to win even more than noobs(why they spent the time to develop the skill in the first place).

 

[Phelan]

I think it depends also on what a player percives as entertaining. Some of us are only entertaint when they can beat the game. They really look for the game mechanics and make clever plans to overcome them. Just look at the discussions in the HOI Forums about "the optimal elite division" or "zerg-tactics". I really dont understand this kind of attitude. For me this is really boring. What I like on computer gaming (but also on boardgames) is the emmersion, the feeling, the roleplay. If I play England in Europa Universalis, I dont try to smart out the French AI to Annex whole of France in just 20 Years. That is boring, I know that I might be capable to do this, if I would try to figure out the weak spots of the AI, but I dont want it. This is no fun for me. It is a bit like playing with your child. I know I can beat it when we play a board game, but I have really fun to play WITH it and not AGAINST it. Thats the same attitude for me with game AI, I like to play with the AI and not AGAINST it. The AI must however be sufficiently capable to do things in a game. They shouldn't do it perfect, but the should be able to do it. The best example at the moment are the narval invasions in HOI3. I dont expect them to do always really good invasions, but in principle they should be able to do a meanigful narval invasion. That is what PI must learn the AI, not perfectionism.

Oh my, this is a really complicated matter...

 

[kierun]

I suggest anyone interested in AI read Artificial Intelligence: A Modern Approach (Prentice Hall Series in Artificial Intelligence) by Stuart Russell and Peter Norvig. Then try coding one for a laugh...

 

[Yaromir]

Some of us are only entertaint when they can beat the game.

I think it is a safe bet that most people want a combination of two things:

1. System which demands application of meaningful effort (That is, game should be Challenging)
-AND-
2. That the player wins.

I think highest degree of satisfaction is achieved when both of those components are present (Game is challenging but you win).

 

[unmerged(75409)]

I think it is a safe bet that most people want a combination of two things:

1. System which demands application of meaningful effort (That is, game should be Challenging)
-AND-
2. That the player wins.

I think highest degree of satisfaction is achieved when both of those components are present (Game is challenging but you win).

I don't mind losing a game, actually. As long as the AI does not single me out for aggression (like AI UK trying on purpose to stomp the player country).

Some of the best moments in Vic 1 came after realizing my country was losing the war against Germany and France (I was playing Greece ) and I needed to give up something to get peace. I got a Bolshevik takeover after that war (I tried to get the Socialists elected but must have chosen the hard left option in the election events once too often o)

Rebuilding a broken country after a lost war can be a good challenge.

Paradox games aren't "won" in the conventional sense anyways...

 

[unmerged(68110)]

I can tolerate the occasional AI brainfart - but cheating simply sucks.

Waging war outside of the event scope of Vicky I or VIP with a great power as the enemy can be a serious pain in the arse, as they are able to convert one bajillion pops into soldiers, recruit them all in a single move and keep a thousand dreadnaughts WITHOUT PAYING UPKEEP.

Seriously.

 

[unmerged(63310)]

I don't mind losing a game, actually. As long as the AI does not single me out for aggression (like AI UK trying on purpose to stomp the player country).

Paradox games aren't "won" in the conventional sense anyways...

With your first point I am not sure if you can really claim AI singled out human player... more that AI singled out weakest victim which is usually the most effective strategy actually so I can't say that is a weak part of the AI.

The 2nd point is absolutely true though and gives developers much flexibility in design opposed to something like HoI3 where winning or losing was more clear cut.

Although as people keep mentioning the sense of the AI getting unfair help to achieve something the human player cannot in a game without clear cut victory conditions might be even more of a problem. IE- when human player triumphs despite AI getting unfair advantage there is some satisfaction for most people. When there is no real triumph and the AI continues getting some obvious advantage it might be more annoying.

Personally I only get annoyed when the AI "cheats" also coincide with something historically ridiculous or extremely unlikely. 500 BB fleets or 1 million man armies in Africa for example.