Higher difficulty level = Smarter AIs?

[Pinstar]

I was listening to the official Civ V podcast the other day where they talked about AIs and mentioned that on higher difficulty levels, they are programmed to be smarter, basically speaking.

Does that mean that they will be removing the raw stat bonuses (Faster tech research, free scouts/settlers) from the AIs on the higher difficulty levels and will rely more on the smarter decisions made by the AIs to give the players a run for their money? Or are they just going to combine the new 'smarter AI' on higher difficulties with the traditional raw AI bonuses that the higher level AIs got in older versions of Civ?

 

[tom2050]

I'm sure AI will still have bonus'; it's not possible to make the AI that smart (with a budget) to take all bonus' out and still have an AI capable of defeating a human player.

 

[Ogrelord]

http://forums.civfanatics.com/showthread.php?t=378410

 

[poncratias]

I am afraid that you can't have the smarter AI of the higher levels but without all those aggravations and bonuses for the AI...

would have loved that...

 

[MkLh]

It doesn't sound a good idea to make AIs in low levels less smart than they could be. Most lower levels players probably think that handicapped smart AI is better than non handicapped stupid AI.

 

[Schuesseled]

I don't have a problem with having different AI difficulty levels, I just wish they would seperate that from the handicap bonus.

 

[SeismoGraf]

Having smart AI would be ideal; handicapped at lower levels, but no or little bonus on the highest difficulty settings. It would also improve MP matches, where a leaving player could be replaced by a comparably competent AI; maybe based on the player's stats or his past success in the ongoing game.

 

[IdleEnergy]

Having smart AI would be ideal; handicapped at lower levels, but no or little bonus on the highest difficulty settings. It would also improve MP matches, where a leaving player could be replaced by a comparably competent AI; maybe based on the player's stats or his past success in the ongoing game.

So the AI should be balanced around beating extremely good civ players straight up without bonuses and then handicapped from there? Sounds almost as expensive as it is entirely unrealistic.

Chess only has 32 pieces, 16 per team with 7 different kinds of pieces total, operating in 64 homogeneous tiles, and computers have only recently been able to beat the best humans in that narrowly defined space. CiV has 18 unique civilizations with variable sized maps, heterogeneous tiles, 72 hierarchical technologies and 56 basic units. There's no way a computer is going to be able to beat a very good human player without bonuses.

 

[SeismoGraf]

So the AI should be balanced around beating extremely good civ players straight up without bonuses and then handicapped from there? Sounds almost as expensive as it is entirely unrealistic.

Chess only has 32 pieces, 16 per team with 7 different kinds of pieces total, operating in 64 homogeneous tiles, and computers have only recently been able to beat the best humans in that narrowly defined space. CiV has 18 unique civilizations with variable sized maps, heterogeneous tiles, 72 hierarchical technologies and 56 basic units. There's no way a computer is going to be able to beat a very good human player without bonuses.

Nobody has done it, but I also think it would take a lot of development effort. Teaching computers how to play Chess has been a subject of academic study since the beginnings. If I remember correctly, Alan Turing conceived a chess program, even before the computer was invented.

Over the decades millions of dollars went into developing Chess AI, mainly public funds. Then came the unthinkable: Deep Blue, "a chess-playing computer developed by IBM that defeated world champion Garry Kasparov in 1997" (Wikipedia).

I am a Chess player and have some knowledge in programming. The strongest PC Chess software currently on the market is assumed to draw its strength from superior heuristics and algorithms. It can be safely said, that 95% of club players will never win a game against this SW, unless there is a bug, of course.

The development of processing power and memory capacity helped Chess AI get stronger. But the major improvements came from algorithms (e.g. Alpha-Beta pruning) and heuristics ("game knowledge").

Have you watched Playing to Lose: AI and "Civilization" (Soren Johnson)?

In this video you can find many answers. As I understand it, the main goal of the developers was to find a good compromise between "good AI" and "fun AI".

I have no doubt, that what happened to Chess players, will one day also happen to Civ players - humans will get surpassed. Regardless of the combinatorial complexity of Civ, that far surpasses games like Chess and Go. Whether in 10, 20 or 50 years, nobody can predict.

Getting steamrolled on settler difficulty by the AI would be a bitter pill to swallow

ps: "There's no way a computer is going to be able to beat a very good human player without bonuses" - that is exactly what the Chess community was saying, in a time, when computers still played Chess like beginners.

 

[mercury529]

Computational power limitations certainly enter into the picture when discussing AI in a game like Civ. Perhaps a particularly ambitious modder will mod an AI enhancement mod that uses the computational power of a botnet .

 

[Zhahz]

In GalCiv2 at lower difficulties the AI isn't fully enabled. You could say it's capable of making mistakes or something. As you scale up the difficulty, the AI is eventually fully enabled. If you ratchet up difficulty enough, the AI is fully enabled and starts getting resource bonuses.

To me if you're a lesser player playing on lower difficulty levels, it would be a lot more interesting to play against an AI that can make mistakes (like a real player). Real players wouldn't square off with production variations.

From what I've read/heard the Civ V AI doesn't really make mistakes at lower levels, it just evaluates things more deeply and does more big picture strategizing at higher levels, which is also more like how some players vary in skill.

 

[better days]

Chess only has 32 pieces, 16 per team with 7 different kinds of pieces total, operating in 64 homogeneous tiles, and computers have only recently been able to beat the best humans in that narrowly defined space. CiV has 18 unique civilizations with variable sized maps, heterogeneous tiles, 72 hierarchical technologies and 56 basic units. There's no way a computer is going to be able to beat a very good human player without bonuses.

Agree with this- there's simply no way that a computer can beat the best Civ players.

So how I would set up the difficulty levels- at the lower levels, instead of giving bonuses to the player, I would make the AI play dumber. If Civ 5 was my fist Civ game, I would most certainly play at the lowest difficulties, and work myself up as I improved. Personally, I would like to know that the computer wasn't playing as smart,as opposed to me starting with extra workers/settlers etc. at the lower levels.

Using the chess analogy again, when you play against a computer and set the difficulty at a low level, it doesn't remove 2 pawns and a knight of the computers, but rather makes it play less optimally. I think this approach would be better for Civ than player bonuses.

However, if a good Civ player was regularly able to beat Civ 5 on the highest level that the AI was capable of, which I think will be very probable, then there are only two options:

1. Go play MP - Not getting into the whole MP vs SP arguement, but a lot of players don't enjoy multiplayer. There needs to be something else to challenge these players. The only option is to...

2. Give the AI bonuses (production, research etc.) at higher difficulties (ideally separated from the AI "brainpower" handicaps, so that the AI recieves no bonuses until it's maxed out its ability). Furthermore, I think that whatever the highest difficulty is (Deity? is that confirmed at all), it should be impossible for the average to good Civ players, and only the very best would even have a chance to beat it. If that means crazy bonuses at the higher difficulties, so be it. I believe Civ 5 needs to be able to offer a challenge to everyone (something, IMHO, Civ 4 did very well).

Just my 2 cents.

 

[KrikkitTwo]

Agree with this- there's simply no way that a computer can beat the best Civ players.

So how I would set up the difficulty levels- at the lower levels, instead of giving bonuses to the player, I would make the AI play dumber. If Civ 5 was my fist Civ game, I would most certainly play at the lowest difficulties, and work myself up as I improved. Personally, I would like to know that the computer wasn't playing as smart,as opposed to me starting with extra workers/settlers etc. at the lower levels.

Using the chess analogy again, when you play against a computer and set the difficulty at a low level, it doesn't remove 2 pawns and a knight of the computers, but rather makes it play less optimally. I think this approach would be better for Civ than player bonuses.

However, if a good Civ player was regularly able to beat Civ 5 on the highest level that the AI was capable of, which I think will be very probable, then there are only two options:

1. Go play MP - Not getting into the whole MP vs SP arguement, but a lot of players don't enjoy multiplayer. There needs to be something else to challenge these players. The only option is to...

2. Give the AI bonuses (production, research etc.) at higher difficulties (ideally separated from the AI "brainpower" handicaps, so that the AI recieves no bonuses until it's maxed out its ability). Furthermore, I think that whatever the highest difficulty is (Deity? is that confirmed at all), it should be impossible for the average to good Civ players, and only the very best would even have a chance to beat it. If that means crazy bonuses at the higher difficulties, so be it. I believe Civ 5 needs to be able to offer a challenge to everyone (something, IMHO, Civ 4 did very well).

Just my 2 cents.

Well Ideally the way I see the game difficulties

At Easiest:
Human bonuses... higher happiness/less maintenance/anti-barb bonuses (make the game more forgiving for easy players)
No AI bonuses
AI has some artificial stupidity

As difficulty increases
FIRST: remove Human bonuses (the human is able to play the game)
SECOND: once Human bonuses are gone, remove AI artificial stupidity
THIRD: after all AI artificial stupidity is gone, give AI bonuses


Now Ideally, they could allow a 'custom Difficulty' level where you
1. Choose AI intelligence
2. Choose General game difficulty
3. Choose AI bonuses

 

[Earthling]

I am a Chess player and have some knowledge in programming. The strongest PC Chess software currently on the market is assumed to draw its strength from superior heuristics and algorithms.

This sounds rather uninformed considering your stated background, but one of the hugest strengths of chess computers is drawing on tons of information about previous games and statistics from memory. Essentially, you'd describe it as having access to books of openings and endgames and other theory so on in the middle of a game.

Chess computers still cannot compare with humans when this advantage isn't present - eg. Fischer random chess. And players still have beaten top computers recently taking advantage of such flaws - playing to openings the computer doesn't have memorized, in other words.

Edit: was just posting other thoughts about civ AI in the other thread though but I figure I should sum up - I would rather not have the AI "dumbed down" at low levels but I probably won't even play those low levels anyway. I do hope the bonuses, diplomacy and so on are improved over civ4 though.

Also, the googletalk you linked is very nice, thanks for doing so. Really explains why Civ IV had such improvements in AI and gameplay there for anyone who hasn't seen it I would also encourage doing so, I really hope Soren Johnson's philosophy and influence shines through to continue in civ5 though.

 

[snapple232]

If Civ was simple enough to be "solved" by an AI (aka able to hold its own against the best human players) it'd be too simple to be worth playing.

 

[Drakan]

Would humans buy a game which AI systematically steamrolled them regardless the difficulty setting? Nah.

We are probably just a few decades away until this happens.

 

[DeepQantas]

I'd rather not cheat to beat an AI.

 

[SeismoGraf]

This sounds rather uninformed considering your stated background, but one of the hugest strengths of chess computers is drawing on tons of information about previous games and statistics from memory. Essentially, you'd describe it as having access to books of openings and endgames and other theory so on in the middle of a game.

Why do you think I'm uniformed? Because I haven't mentioned opening books? There is no doubt, that they increase the playing strength significantly. Removing the opening book from a Chess program makes it weaker (say -200 ELO); remove Alpha-Beta pruning and replace it with Minimax, it will hurt terribly (say -500 ELO).

Endgame databases (large collections of solved positions with few stones left) can make a decisive difference in a single game, but be completely irrelevant in others. In 99% of games the average human is lost, even before queens are exchanged. In computer vs. computer, analysis or correspondence play, these databases are more important.

Without knowing your background, I think it makes sense to focus the discussion of Chess AI to areas where it is connected to Civ AI.

Chess computers still cannot compare with humans when this advantage isn't present - eg. Fischer random chess. And players still have beaten top computers recently taking advantage of such flaws - playing to openings the computer doesn't have memorized, in other words.

An expert or world-class player winning 1 game out of a whole bunch by exploiting flaws can't be considered a big success, can it? Show me a player that can consistently achieve a "mere" 33% score in Chess960 against the best software. There appears to be none on this planet.

Chess software demolishes humans, even when not having access to opening books. When playing standard Chess, the human can still rely on his memory of openings, thus usually last a bit longer, before his position collapses.

Have you ever played a game of Chess960 against a strong chess program? The human player can't rely on his memory either and has to think deeply beginning with the initial position.

In 2007 there was an 8 game match, Rybka vs GM Joel Benjamin, with a human grandmaster playing against the computer with a handicap of 1 pawn. To illustrate, I would compare it with the advantage of starting a Civ game with 1 Worker for free. It is no guarantee that you'll win, but your odds have increased by a lot. The match was won 4,5:3,5 by the computer. Playing without handicap was judged uninteresting, because the winner would have been known in advance.

If Civ was simple enough to be "solved" by an AI (aka able to hold its own against the best human players) it'd be too simple to be worth playing.

Civ can't be "solved" like Checkers. Chess can in theory be solved, but e.g. Poker can't. One reason is, that Civ and Poker are not games of perfect information.

 

[Souron]

So the AI should be balanced around beating extremely good civ players straight up without bonuses and then handicapped from there? Sounds almost as expensive as it is entirely unrealistic.

Chess only has 32 pieces, 16 per team with 7 different kinds of pieces total, operating in 64 homogeneous tiles, and computers have only recently been able to beat the best humans in that narrowly defined space. CiV has 18 unique civilizations with variable sized maps, heterogeneous tiles, 72 hierarchical technologies and 56 basic units. There's no way a computer is going to be able to beat a very good human player without bonuses.

That's not a fair comparison. Chess programs don't think like humans. The Heuristics that chess programs use more closely model what programmers think a perfect player would do, then what actually goes through the mind of a grandmaster. This can be exposed though various situations that computers do poorly in, though the memory or processing time can overcome any deficiency.

Chess programs work by considering a large number of possible moves and evaluating the benefit of the outcome. Is that how you play civ? It would not be a good way to design Civ AI either.

 

[Souron]

Would humans buy a game which AI systematically steamrolled them regardless the difficulty setting? Nah.

We are probably just a few decades away until this happens.

The trick to AI steamrolling is to make the game simpler. For instance, AI is perfect at tic-tak-toe. Humans can be perfect too, so that's not so impressive. But by the same token, AI can be perfect at Checkers, which most humans certainly aren't.