It's fine for the AI to have some knowledge cheats if it's to speed up the turns and results in reasonable game play.
Yup, and having it cheat through knowing more seems like a more reasonable proposition both in terms of player acceptance (having them know more to make up for their rigid algorithms is something I see people more willing to accept than having them receive stat-based bonuses that cause them to be flatout better off in any situation where their algorithms actually approach perfect behavior) and in terms of driving AI designers to make better AIs (if the AI knows everything it possibly can about the game and can still be beaten easily by the human, you have made an uncompetitive AI)
I'm still unconvinced that an AI can be built that gives the feel of playing against a tough opponent with the 1 upt system.
If it were consistent 1UPT (so air units wouldn't stack, and civilian units like Great Generals and embarked units could not occupy the same tiles as military units) and nothing else (so no melee-units-take-damage-on-attack-but-ranged-don't and no insta-heal promotion), it could definitely be possible, as proven by the AIs in various wargames people have brought up in this thread. Problems usually arise when the AI is supposed to build and send units over long distances with 1UPT: it's a problem even with unit stacking (since the AI needs to have a sense of where units are needed and what type are needed), but it becomes exponentially more difficult when units are made so valuable via Civ5's many additions (1UPT, unit resource usage, combat that doesn't always result in one of the participants dying or withdrawing, UU's that have promotions that carry over when upgrading). For example, having the AI create a Longbowman vs. a Crossbowman in Civ4 isn't that much of a problem because it'll be sitting in a stack of other units anyway and isn't a specialized unit like a siege weapon (Longbows in Civ5 were better on defense and had more first strikes, but Crossbows had slightly higher combat strength and had a bonus vs. melee units and were unlocked one or two techs after Longbows), but having it create a Swordsman vs. a Composite Bowman in Civ5 is a huge, huge difference, not just because of how CompBows can ranged attack and Swordsmen cannot, but also how Swordsmen cost 1 iron to maintain and must be adjacent to its attack target (whereas CompBows can fire over another, blocker unit).
Soren Johnson (Civ 4 designer and AI programmer) had a interesting talk on AI's
https://www.youtube.com/watch?v=IJcuQQ1eWWI&feature=sub
Sometimes you have to change the game to allow the AI to be effective. Soren talks about several changes made to Civ IV that enabled the AI to be effective. Apparently this was not done for Civ V
Thanks for linking that talk, I hadn't seen it before.
I respect Soren Johnson quite a lot, especially his work on Civ4 and his newest game, Offworld Trading Company (which I absolutely love).
However, he did make one very crucial, bad assumption in that talk: he assumed that in a game like Civ5, a player can only have fun if they win. If that were the case, then yes, an AI that plays to just barely lose compared to the player is the way to go if you want an AI that gives the player the best experience. However, I will argue that there are cases when the AI is too challenging for the player, but the player can still have fun; the biggest caveat is that all of these cases hinge on a good diplomacy AI who can be talked to and dealt with as easily as a human player. I'm not just talking about backstabbing and 2-on-1 gangups, but I'm also talking about large coalitions, being able to put differences aside to focus on a common [AI] enemy, waging a cold war... and more importantly, it needs to be able to bring the human player in on these things. Being a kingmaker can be fun, especially if AI personalities have some persistence (ie. if you helped Ghandi to win in the last game, he'll be more fond of you in the next game you play with him) and/or if future civ games allow for multiple victors. In FFA games, human players only stop having fun if they don't feel like they are doing anything important in the game, and a competitive AI that wipes the floor with the human player can still allow them to act as a kingmaker between an AI who backstabbed them and the underdog AI who is trying to beat the other AI by helping the human player (an enemy of their enemy) out. It can also make the human player feel like they've won a personal victory even if they do not win the actual game: examples include generating a huge profit by playing both sides of two, much stronger AIs engaged in a cold war, or annihilating the AI that almost destroyed them earlier even if their strong AI ally will still win a spaceship victory in the next 5 turns, or denying an "" AI World Congress hosting by voting for another AI.
So yeah, in an FFA game, an AI can play to win and still make the human player feel like they are having fun (as well as giving them a lot of interesting stories to tell about exploiting AI relations that feel like they're exploiting human players instead of poor AI coding). Things are much different when teams are static though.