Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[forsaken1111]

Finally got around to playing with the newest version, but I think I'm also already done with it. While I cautiously approve of how they've handled populations and planets, and I'm glad they finally removed that stupid leader cap, overall I feel like they just massively slowed down the game to little other benefit. Research proceeds at a crawl, while fleet and station caps are absurdly low so everyone's maximum fleet size is minuscule. Until now, 60 years in, I've seen exactly one war in the western half of the galaxy and there's no chance of a war anytime in the near future because neither I nor any of the AIs around me can muster enough fleet to make one worthwhile.

Yeah I have to disagree too, unless you are still very early in the game? I've had no trouble doing early wars and stacking tech, to the point that I usually out-tech most nearby AI players to an absurd degree. Also, your fleet cap is more of a soft cap. You can exceed it, and SHOULD exceed it if gearing up for war. You're probably going to lose a bunch of those extra ships anyway so you won't be paying the maintenance for long. Stock up on alloys and credits beforehand so you can finance the war then build way over your cap. I sometimes go 50-75% over to ensure I have more firepower than the opponent.

Station caps are a bit low, just be judicious about where you place one. You don't need stations everywhere, just at chokepoints or at positions where one station can cover several trade resources. Once you get larger stations a single trade station can cover many systems so you can start removing old ones and making your trade network more efficient.

[Radsoc]

The AI is kinda meh. I still think they need to put resources into experimental neural network training and so on, where you set difficulty by generation number. The big challenge with the game are the events though and that's ok, because it means that competing with the AI is not enough (they are just there to fill the universe and to obstruct or help you), you have to reach a certain threshold of development to survive.

Performance wise there is a difference between my i7 gen desktop and my i5 gen+1 laptop, where the end game is playable but not at any kind of sensible pace. I still don't get why people recommend i5. They just need to be replaced more often.

Don't know if you read about the utopian abundance "exploit" where you can use that mechanics to basically blow away all competition. It's hilarious.

[EnigmaticHat]

That would be a waste of resources... neural nets operate best when they have set inputs and outputs.  The number of considerations the Stellaris AI has to think about, and the number of orders it has to give, vary a lot based on game state.  Neural net based game AIs literally just play the game as a human does; their inputs are the color values of the pixels on the screen and the outputs are the keyboard + mouse inputs.  Very static.  Since you can't render the game 20 times a frame for 20 different AIs, you'd need it to read the game data directly.  The "graph" formed by hyperlanes and planets is inherently unfriendly towards neural nets because its shape and size are different every time.  And that's just scratching the surface of the problems you'd have.

[Radsoc]

Well, yes, the implementation in some games right now is that you take what you see on screen, but it doesn't have to be like that. Neural nets can be used and trained in countless ways. There would be problems, but nothing that can't be overcome. I think that normal algorithms, non-stochastic, non-evolutionary and non-neural, are the best for most commercial tasks, because they are easy, predictable, and get a minimum done right, but they are a lazy way out for AI.

[E. Albright]

Evolutionary AI is only as good as the feature set you provide it, and feature selection is not a trivial task. "Give the AI all the data it can possibly have, plus state-by-state input sets from successful human-controlled games, and let it try to hash out feature selection evolutionarily" doesn't necessarily solve the problem either, because some of the inputs humans are considering are metadata. You can argue that old-school rationalist AI is a cop-out, but it's not a given that empiricist AI is going to do any better - it's entirely possible it could do worse.

[Paul]

Just teaching the AI some basic patterns for efficient colony management and having it use advantageous things that the player uses (like encourage growth decision) would go a long way. Right now they grow really slowly and build so sub-optimally that they fall behind even with a huge handicap with the highest difficulty. I've looked at some of the AI empires part way through games and found them suffering from special resource and mineral and food shortages and totally crippled, while their worlds that should be producing food and minerals are short on workers due to everyone being promoted to specialists running the big upgraded buildings which are causing the special resource shortages.

[Radsoc]

The thing is that Stellaris is a numbers based game, with scores that could be used as a base for "fitness" and what not. A hands off approach would have the AI select things at random, and then check against the score at specific times. Most decisions would be detrimental and filtered out. Then you save and build winning sequences (not necessarily only the best) with probability weights for future actions. Of course there are huge problems and it's far from trivial, but leaving mechanical AI behind is the only way forward if we ever want opponents that act like humans or better. Winning neural nets can be evolved in similar ways where difficulty is set by generation number.

[E. Albright]

Intermediate score-based analysis seems like a very fraught path to pursue when a game has as much potentially far-reaching randomness as Stellaris. This sort of approach could easily lead to AIs who do reasonably well most of the time but completely fall apart under particular circumstances. More generally, having a good score at a particular moment is not perforce a good predictor of future score; not all situations producing good score values are equal.

I'm not saying it's a hopeless idea, but it's definitely not a magic bullet.

[Loud Whispers]

One of the cool things about Age of Mythology was some experimentation done with giving AI personalities. Besides obvious AI improvements to executing strategies, they found you could make more human-like AI by giving them a certain level of personality - AIs that overestimated the strengths of their opponents or underestimated the strengths of their opponents being one such factor. It's rather alike CK2's personality system, where cowardly, brave, zealous, greedy or mad characters were more or less likely to go to war at certain times

[EuchreJack]

Just teaching the AI some basic patterns for efficient colony management and having it use advantageous things that the player uses (like encourage growth decision) would go a long way. Right now they grow really slowly and build so sub-optimally that they fall behind even with a huge handicap with the highest difficulty. I've looked at some of the AI empires part way through games and found them suffering from special resource and mineral and food shortages and totally crippled, while their worlds that should be producing food and minerals are short on workers due to everyone being promoted to specialists running the big upgraded buildings which are causing the special resource shortages.

Having played a bit, it is possible that resource shortages could be a strategy based upon the market, where its usually cheaper to buy food and minerals and sell alloys.  Instead of the AI being "broken", it might have just adapted to the implementation of the market.

[Radsoc]

Intermediate score-based analysis seems like a very fraught path to pursue when a game has as much potentially far-reaching randomness as Stellaris. This sort of approach could easily lead to AIs who do reasonably well most of the time but completely fall apart under particular circumstances. More generally, having a good score at a particular moment is not perforce a good predictor of future score; not all situations producing good score values are equal.

I'm not saying it's a hopeless idea, but it's definitely not a magic bullet.

That would be a place for the neural net. You generate actions sequences on the go (and from memory), but the neural net finds patterns in when to employ them. This would of course require Paradox to fund things that may not be commercially viable within reasonable time frames (for them) and they could very well survive and keep rolling without improving AI that way, but once more games employ new AI, that kind of companies would only survive on licensed thematic trademarks or MP features (which already ruin games due to too conservative near-symmetrical balance). The total war series is another example there.

[Telgin]

One of the cool things about Age of Mythology was some experimentation done with giving AI personalities. Besides obvious AI improvements to executing strategies, they found you could make more human-like AI by giving them a certain level of personality - AIs that overestimated the strengths of their opponents or underestimated the strengths of their opponents being one such factor. It's rather alike CK2's personality system, where cowardly, brave, zealous, greedy or mad characters were more or less likely to go to war at certain times

Stellaris does this too, to a point, doesn't it?  I've never really noticed much difference though, since I think most of the difference manifests as opinion modifiers, but a few personalities declare war more readily.

[forsaken1111]

I couldn't even imagine how you'd begin trying to train an AI on a problem set where there may be hours between choice and consequence. Even humans who are good at games like these don't always understand what they did wrong to lead to a defeat. Google might have a crack at it, but I doubt Paradox even has the resources to attempt it.

[E. Albright]

That would be a place for the neural net.

You ARE treating this as a magic bullet, though. Alphabet used neural nets to beat a human Go master on a full board w/o handicaps in 2015. In Go. Go is much simpler than something like this both in terms of feature space and possible actions, but it took Alphabet throwing resources at this to get to where an AI could beat a master a couple of years ago.

We're not there yet, and I'd not hold your breath. As forsaken1111 points out, this is a VERY messy, poorly-defined problem for an AI to solve.

[EnigmaticHat]

The neural nets used for DOTA 2 and SC2 resulted in these sort of trees of different iterations of the bots, the branches being pruned by eliminating bots that would under-perform against other bots.  Most of the branches had a specific strategies they would go for.  Like there was one version of the Starcraft 2 bot that would never build anything other than stalkers and was just impossibly good at microing them.  The bots also floundered at anything that didn't have a direct, immediate payoff towards something they cared about; for example the DOTA 2 bot couldn't understand sentry wards and would place them under enemy towers to tank exactly 2 hits every time.

While those bots were promising for the future of AI, from a game dev standpoint they're not appealing.  The game dev's goal isn't to BEAT the player, its provide them with a convincing and satisfying opponent.  As mediocre as Paradox's AIs might be, they're better than what a neural net would currently produce.