Zurai posted:

Nice straw man. Try actually addressing the point I was making instead of making up a different point I never mentioned.

EDIT: Seriously, the permutations for 12 units, each of which can make 12 different possible moves/attacks, is half a billion possible sequences of moves (and sequencing matters in 1UPT). Most Civilization game states past the very early game are much more complicated than that.

A Wizard of Goatse posted:

OK, you can just repeat the same thing for every ineffective strategy until you've got the half-dozen or so that actually produce a serious threat to a comparable opponent? There are literally not that many ways to play Civilization well, there's like one or two places you can put your spearman where it'll be relevant in the next half-dozen turns and if you're not doing that I don't care. I'm sure your secret kung fu move of canceling production and moving units at random and pursuing the not immediately useful research paths is extremely powerful but there's no way to represent that that I don't think you'll take as "strawmanning".

Zurai posted:

This is relatively easy for a human to figure out and pretty much impossible for a computer to figure out. Computers don't have intuition. They can't say, "Oh, this move doesn't actually matter" without pursuing a line examining whether it matters.

Also, good job playing the victim when you're the one who ignored the entire point to invent a moronicly-weak scenario to argue against.

Zurai posted:

Sure they are. You have dozens of cities, each of which have dozens of things they can do, including canceling the thing they were already doing to do something else. You have dozens of military units, each of which can move and attack in dozens of different ways on any given turn. Just the permutations on those with different sequencing is best expressed in scientific notation. That's without counting things like governments, techs, workers, etc.

Unlike Chess, the Civ AI absolutely cannot crunch through every possible reaction to its moves. There simply is no way to do it.

GlyphGryph posted:

Actually, "intuition" is probably the best description for what machine learning does. It absolutely does not, in most well developed system, pursue any "lines" when deciding whether something is a good move or not, it's usually pattern matching and state comparison. "These kinds of moves generally produce good outcomes in similar situations, so I'm going to do one of those"

GlyphGryph posted:

Hey, something I'm actually knowledgeable about! Machine learning is very very good at the sort of complexity you're describing, and no you don't need a human opponent or an existing AI to train them. 4x games would be a remarkably easy problem for adversarial networks to solve in fairly robust ways with modern design approaches. It doesn't even matter that games are particularly long (and they wouldn't be all that long for an AI) because any decent training network isn't going to be playing full games anyway.

Absolutely no one developing 4xs right now is likely to have the skill to do any of that, nor do they have any real incentive to gain or develop or them, because that involves a lot of work and technical know-how, but it wouldn't be particularly difficult to train an AI to play a 4x game well enough to completely outshine any human player, since they are complex enough you don't need to get anywhere near "optimal" play, you just need heuristics slightly better than the ones human use.

But who actually wants that? Do players care enough to make it worth the money to hire the talent or buy the hardware needed? Hell no, and most players would complain about the AI "cheating" if you bothered. This is video game development! If you can take a shortcut to avoid doing work the player can't see, that's what you're gonna do!

Ardryn posted:

There's a video floating around youtube of all the projects back to back and if you take them with the tech quotes early on you see a lot of Miriam, but then she disappears midway or so through it and then re-appears near some endgame stuff like the self-aware colony _{(We Must Dissent)} and the teleportation gate wondering if it remembers to transfer the soul as well. Just all around great writing that really made a ton of SMAC work despite the faults, technical restrictions, and gameplay weirdnesses.

I also love one early game quote that says "Einstein said that God does not play dice." and then a much later quote says "Not only does God play dice, the dice are loaded."

Rappaport posted:

It's a fun thought experiment, and thanks for this post. I personally would not want to play against a super-AI. At least for me, the fun in 4X's is that there's a definite sense of progression, and your main "enemy" is the economic system and trying to snow-ball yourself to victory. It's also fun to hate on Miriam or the Sakkra, and trying to deal with their sub-optimal bullshit, but I see it more as keeping their filthy hands away from my stuff, which I need to make more stuff, and on and on. If the AI could just merk me every time no matter what I did, what's the actual game left there?

Or in other words, the real Skinner box in a 4X, in my view, is getting new techs, colonies, making more pixel mens, etc., rather than killing the AI's pixel mens. This is why these games try to have different winning conditions than "you must murder everybody", because the process of doing that is usually very tedious and uninteresting near the end of the process.

victrix posted:

personality based ai with completely transparent behavior is a much better system even if it results in stupid or sub optimal decisions, because you can play around them, and mixing them in one game creates interesting and fun situations (sup ai war)

but this assumes you get your game to a place where the ai is competent in the first place by launch which, well *gestures at genre*

Demiurge4 posted:

Coming in to play the black sheep because I picked up Distant Worlds 2 the other day and I hate that I can't micro hard enough!

I went into the pre-ftl start so I could goof around and meticulously manage my expansion down to the nitty gritty details but the game actively works against you on this through the auto refit mechanic. I had designed a small mining station and I figured that would go on deposits of rare materials because those are used in small amounts so I could just have a small stockpile, then I unlocked the medium station and made a design for that to go on gas giants and asteroid fields with lots of basic materials and POOF, all my mining stations everywhere auto refit themselves to the medium station design.

I ended up tuning it so that my designs will only upgrade into their own hull type but my next problem is that the construction ship UI doesn't allow for this, it will always highlight the biggest design on the UI and it's killing me. I want to play a tall empire with minimal sprawl that I can meticulously micro manage the but game is fighting me on it.

Squiggle posted:

Maybe I'm missing something that's not included in the options, but - if you go to your Empire menu and view Policy Settings, does running down the side mashing "Manual" on every option not give you full control?

EDIT sorry for the quality on this, it's from a remote window back to my desktop at home, but:


There's actually a Preset Configuration for "None" automation. You can't really tell in the shot, but it scrolls and there are a LOT of options here - is there more being automated beyond these settings? I think some of the things you're wanting a direct hand over are under the "Construction" section.

I do believe the private sector stations and ships will always self upgrade though. Free enterprise, babyyy

GlyphGryph posted:

Hey, something I'm actually knowledgeable about! Machine learning is very very good at the sort of complexity you're describing, and no you don't need a human opponent or an existing AI to train them. 4x games would be a remarkably easy problem for adversarial networks to solve in fairly robust ways with modern design approaches.

GlyphGryph posted:

Also, good luck getting someone who knows anything meaningful about machine learning to work at a game company at all in the current environment, hah.

Caconym posted:

Pop management is still my main gripe, I want my planet pop policies to apply to newly settled planets dammit, don't fill up my one and only boskaran volcano planet with humans, I need my super troopers.

Subjunctive posted:

I’m surprised to see that nobody seems to have done anything with a 4X after AlphaStar solved SC2. I wonder why that hasn’t happened.

Jabor posted:

Anyway it turns out that even mid-level Go players can pretty reliably beat top AIs by exploiting specific weaknesses in how they evaluate the position. If you tried to machine-learn up your own 4X AI, it would almost certainly have a similar blind spot - and now not only is there a single "optimal" way to play, you as the designer have absolutely no control over what it's going to be.

Jabor posted:

Anyway it turns out that even mid-level Go players can pretty reliably beat top AIs by exploiting specific weaknesses in how they evaluate the position. If you tried to machine-learn up your own 4X AI, it would almost certainly have a similar blind spot - and now not only is there a single "optimal" way to play, you as the designer have absolutely no control over what it's going to be.

Mzbundifund posted:

What are these weaknesses? I thought AlphaZero could beat the best Go players in the world every time.

Squiggle posted:

Maybe I'm missing something that's not included in the options, but - if you go to your Empire menu and view Policy Settings, does running down the side mashing "Manual" on every option not give you full control?

EDIT sorry for the quality on this, it's from a remote window back to my desktop at home, but:


There's actually a Preset Configuration for "None" automation. You can't really tell in the shot, but it scrolls and there are a LOT of options here - is there more being automated beyond these settings? I think some of the things you're wanting a direct hand over are under the "Construction" section.

I do believe the private sector stations and ships will always self upgrade though. Free enterprise, babyyy

FrancisFukyomama posted:

Dont panzer general type games usually get away with it by having pre placed units on static maps? I imagine having to make a tactical AI for a procgen map for units that are actively produced becomes way tougher

Rappaport posted:

It's a fun thought experiment, and thanks for this post. I personally would not want to play against a super-AI. At least for me, the fun in 4X's is that there's a definite sense of progression, and your main "enemy" is the economic system and trying to snow-ball yourself to victory. It's also fun to hate on Miriam or the Sakkra, and trying to deal with their sub-optimal bullshit, but I see it more as keeping their filthy hands away from my stuff, which I need to make more stuff, and on and on. If the AI could just merk me every time no matter what I did, what's the actual game left there?

Or in other words, the real Skinner box in a 4X, in my view, is getting new techs, colonies, making more pixel mens, etc., rather than killing the AI's pixel mens. This is why these games try to have different winning conditions than "you must murder everybody", because the process of doing that is usually very tedious and uninteresting near the end of the process.

Jeb Bush 2012 posted:

I think the whole "ai can't handle 1upt" thing tends to be exaggerated because of civ V/VI having much worse AI than civ IV across the board (not just in combat)

I haven't played a ton of it but the old world ai seems to handle 1upt combat fine

Jabor posted:

https://goattack.far.ai/game-analysis#contents

Essentially, the AI misevaluates a situation that is very rare in "normal" play (a group that comes around and loops back on itself), but isn't difficult to set up by an opponent looking to exploit that weakness.

GlyphGryph posted:

TheDeadlyShoe posted:

In Civ the hardest thing to do is making attacks on cities. A player will queue up their attack force just out of range so they can rush as many units as possible into the assault at the same time. The Civ AI is real bad at that so it tends to suffer defeat in detail. In Old World its much easier for the AI because it can double move up on you.

KOGAHAZAN!! posted:

Well, I think something that's interesting to consider in this context is that it's a very common refrain among players of Civ VI that they find barbs to be more of a challenge than the other civs, and while a large part of that is going to be down to the way that they will rush you down in the first few turns of every game, a lot of it is also that the barbarian AI is more willing to throw away units on the offence that the regular AI isn't. The tests Gort does or did with every patch back this up: the AI fails to take cities, even unresisted, because it will not commit to an assault.

I have a strong suspicion that the AI in that game would look a lot more capable militarily if it just weighted unit preservation a little less.

e: I'll also note that, like, I don't think I've seen the AI have trouble moving units around? It has difficulty keeping up economically- so it has fewer, worse units- and doesn't do as much damage as it could do with what it does have, but actually getting around the place hasn't seemed like a problem for it IME.

my dad posted:

We'll see what Age of Wonders 4 is like at the very least.