I was wondering for a long time:
wouldn't it be possible to create an ML-based AI that could parse CIV games and emulate/imitate human-style play? It would have to be broken down to the respective difficulty levels a human is playing and I guess it would require a lot of input metrics. But people are playing CIV anyway all the time, so there is definitely no lack of data.
That way, CIV could move away from giving the AI boni that it often doesn't know to exploit properly, and we could potentially finally have an interesting competition with the AI.
yes, I would say it's probably possible now but not yet practical. It's really the forefront of AI research and only archivable with a team of top researchers and a lot of computing resources (you need to use a lot of self-play, so AI vs AI). But we still make a lot of progress every year, give it a few years and I would bet that it starts to get practical for game-developers to use machine learning.
The problem with machine learning for game AI is how expensive it is computationally. Forget about using the GPU for it. Maybe if TPUs became more common.
I think it should be possible to make, not a good one, but certainly a much better one, with a combination of various classical game tree search techniques, it's just that AI doesn't seem to have been a huge priority for the developers. There are even mods for Civ 5 that drastically improve the vanilla AI just through tweaking various behaviours.
I think those big models all had visual input and therefore understand what they're seeing. You don't really need this for game AIs, it could make those models drastically smaller.
Maybe, but you'll still be pretty damn limited in terms of access to the GPU. Games generally tend to push the graphics to the limit. You could make a graphically very simple game and put most of your GPU compute into AI. I'd be interested in seeing a space like that explored for sure.
For graphics intensive games though, you either want something that can get by on the CPU or in future maybe TPUs will become common enough in consumer PCs so you could offload your AI onto that and still keep the GPU for fancy graphics.
AI for these types of games is really hard to balance. A lot of players have complained about the weakness of AI in all of the Civ games as far back as I can remember. The problem is that people don’t really want an AI that plays to win. For one thing, it means the AI leaders will be extremely tight and shrewd traders that you can never take advantage of in negotiations. It also means AIs will sign deals with you to size up your military situation and then just wipe you out in a surprise attack as soon as possible.
It means the computer players will no longer role play as historical leaders but instead behave like a bunch of ruthless, psychopathic, cyborg Hitler clones. Most of the historically-informed trade and diplomacy options will be useless because the AIs will just betray you at the worst moment. In essence, they’ll play like human players and honestly, Civ games are just not designed for that. There are much better games out there (mainly board games but also RTSes) for playing against humans who play to win.
I would much prefer all of that over just giving them tons of economic bonuses I can never keep up with. (I'm currently playing Civ 6 on Immortal, where it's simply impossible to keep up at first, and I miss out on all of the wonders, but I still win in the end game simply because the AI is stupid.)
I remember Brad Wardell claiming Galactic Civilizations would have the best AI possible, and it was certainly decent, but players still identified plenty of situations where the AI was being stupid. Brad said he did that intentionally because doing otherwise would be too mean. Thing is, hardcore players did use all of those tricks. At the highest difficulty level, the AI should too.
Use the reasonable, nice, historical AI for the middle difficulty levels, but allow players to choose AI that is just desperate to win.
Since we're talking about freeciv, civ2 on deity is pretty easy if you know what you're doing. It's not even a forced way of playing, you can decide to win with just one city, or you can decide to spam cities with less than 3 population to avoid population penalties, and there's many other strategies.
If you don't know the game well enough to win at the highest difficulty, it is not degrading to play at lower difficulties to learn the ropes while still enjoying the game.
Watching a video of me playing Civ 6 would be incredibly boring. I take a long time, and most of the game I do nothing but desperately figure out how to keep up, and in the end I win with a boring Diplomatic Victory, unless someone else beats me to the Science Victory (other victories don't seem to be possible; except Domination, I suppose). They're not the most exciting games, but they're hard, mostly because it's impossible to keep up in science and wonders.
For my current game, I took a step back to Emperor level, though. I wanted to try an early game Domination Victory, chose Alexander the Great on Pangea and Epic speed, and it's just way too easy (except for holding on to conquered cities during a Dark Age).
I think I'm going to try the same thing on Immortal next. Maybe even slower speed so we spend more time in the lower tech levels where all Alexander's advantages are. And I need to start spying earlier; stealing tech boosts helps a lot in keeping up. I'm still not quite decided whether it's better to stick to a regular age all the time or flip between dark age and heroic age; Heroic Age is awesome, but hard to guarantee, and holding on to conquered cities is very hard during a dark age (I just learned that the Bread & Games project exists primarily for that situation; it helps a lot to have an entertainment district near your front lines).
Of course, most of that just comes down to the game itself having rules that don't actually line up with the intended (for most people) gameplay experience.
Compare to Crusader Kings 3, which is nominally in the same general genre, but has rules that actively push you and the AIs into playing 'in character', because if you just ignore all the secondary stuff your attempts at conquest will almost certainly fall apart.
AI doesn't have to only try to win to be good. It has to give the perception of competition. Perhaps just scaling behind the scenes based on how the player is progressing
I get frustrated with games that scale things with me. I want to be able to feel myself getting better, having my enemies scale with me on any axis really takes away from that feeling that I'm progressing.
I say this not because I think it's an especially noteworthy or important objection, but to echo GP's point, that it's very hard to find AI that suits everyone, and it's not just a matter of difficulty.
Exactly. I do want AI to scale with me, but by me selecting better AI when I get better. That's how I measure my progress: by having to select better AI. That makes my progress transparent.
And by having different AI at different difficulty levels, I think you actually can have AI that suits everyone. But I don't think most game companies like developing several separate versions of AI. They just want one that's superficially "good enough".
I recall seeing Sid Meier say in a talk that he liked to scale his AI to give players the sense of being on a knife's edge and then scaling it back near the end of a game to give a feeling of victory over impossible odds.
This always encourages a human strategy where the optimal behavior is to sandbag in whatever way the AI is measuring your progress, and then slingshot past them once the opportunity arises. Not to say that can't be fun for some people, but it's definitely not a silver bullet.
At leadt in Civ 6, at difficulty levels 6 and above, you can actually loose to the AI. I agree so, that in those cases winning means behaving like a cyborg Hitler clone. There are only so many ways you can hit the leading AI player down on culture or science victories.
I recall one game, my friend and I were playing the strongest AIs. We were closing in on the end of the game, and one pesky AI was very close to a science victory.
My friend and I, both launched a incredible min-max campaign against these guys, using nukes, and whatever fast units we had to conquer and raze cities that had spaceports to stop them. We managed this very quickly, it was like my own desert storm in terms of how quickly and well it went for me.
Then some random AI snagged a diplo victory and I was livid. I disable diplo victories when I play now.
My "current" game is a marathon going for a point victory. It is running for almost a year now, and I still have 30 odd rounds left, sometimes a round, without any real action going on, takes uo to 10 minutes to calculate. So I am in for quite a bit more game time.
In this particular case that isn't the concern. Hitler has a single excellent strategic insight at the start of the war (movement wins, "Blitzkrieg"), but one reason the Nazis lose is that his later strategies don't work. Hitler really wanted to do Sealion (German ground invasion of England) after it became apparent Britain under Churchill will not surrender, but it's just not practical and his general staff persuaded him not to smash his face against that particular wall.
Sealion would be like a Civ AI repeatedly moving a just barely sufficient army from the French coast towards England, each time losing some to the water because they don't have good enough boats yet and each time discovering they can't be reinforced because that's too much movement - so the defenders have an advantage, and the advances all fail. The opponent is bloodied, but loses no territory while the AI's best troops are destroyed for no gain.
A Hitler clone, but perhaps not in the way you meant.
The real Hitler wasn’t a cyborg, to my knowledge. If he were, we’d be speaking German right now. Okay, maybe not, but my point was only to use Hitler as an example that the hypothetical strong AI opponents aren’t to be trusted. The cyborg part is what makes them brilliant strategists, having instant access to computing power to simulate battles.
In Freeciv you don't have total information right?
With imperfect information luck is a factor. An AI which focuses on luck outcomes might be more interesting to play against actually. Why have 100% chance to take exactly one of five enemy cities when you could have 20% chance for each tried in parallel. Less than 1% of the time you crush the opponent with all five defeats, much more often (but still less than half the time) you get none of them, but it always feels like you're a gambler this way.
