Just make you own food ? İt's not even good to begin with from looking at the reviews atleast it's not fume or some other thing that literally worse that it's counterparts and also much more unhealthy (plus it's very expensive)
@@Just_Nice07 sponsors aside, ofc making your own food is going to be better, but this is going to beat most takeout price and quality wise, and much like most things in life the premium is for the convenience
@@ls2260 When it comes to micromanagement, AI can pull of "quantity has a quality of it's own." By sending a small stack to burn down player's countryside, the amount of attention and effort requires to take it down (or at least chase it away) can feel disproportionate, if you also have to manage to actual/proper warfront as well. Of course this is probably a real strategy that has been used to distract/exhaust the enemy. Difference is that Ottomans never sent an army to do that to the Russian Far East.
The thing that makes CK's AI 'good' is that usually, when I am questioning a country's decisions or impressed by how well it's doing, I just look at the leader's traits and go 'Ohhhh, everything makes sense now'.
It’s also funny how absurd it can be. You can challenge a crippled old man with a prowess of -2 and he’ll accept because you killed his son last July. Despite the fact that you Will absolutely kill him. Likewise, you will invite shy characters to a party and they’ll say no.
Allied AI: Runs around in circles with 10 seperate 4k stacks that get stack whiped Enemy AI: Marches around in 40k hivemind death stacks that cross siberia to siege down a human player
Or when you're playing as an indigenous tribe on the West Coast of the US, and suddenly Portugal is coming in 1528 and they already own all of New England and Rio de Prata.
One time when I was playing as Poland, I called Great Britain into a war. They combined their armies into a 120k stack, moved it to London, and spent the entire war dying to attrition.
Classic Stellaris experience: Conquer an AI empire. Spend the next 2 hours paused because you're essentially redesigning every single planet they have.
For context i make or help with stellaris mod making been at it for +3 years(played since launch). Especially AI integration. I felt obligated to explain why the Stellaris AI does what it does. In short, the problems you mention are coming from one area which is there 'Economy plan'. This is what governs how big a budget is for what, when and where it does what. The catch is, a lot of that is based on if resource below/above [number], then give weight increase to secondary resources(Alloys, Consumer goods, Research or military ships). This means it can only make stuff when it needs or has a surplus of said item. It is by design incapable of planning ahead, only ever reacting to its own bottom line. Which means it can be VERY responsive, but has bad specialization as a result. This combined with the fact that the AI is incapable of planning in correlation with its Research, makes specialization always worse when compared to a player. The funny thing about the Ai is, when you get the debug open and see what it wants. It makes sense what it wants. It is not wrong, it just cannot plan. Which makes it all the more tragic in my eyes. Like it knows it needs more minerals, but it cannot remember to have a dedicated mineral world. So it just dumps it somewhere with a research building because there was an unemployed pop on that planet.
Apparently an older version of the AI built of sorts of planet templates that would guide them towards specialization. A bit like people (or at least I) do. The problem was that the templating was too rigid, and the AI would just conk out of it couldn’t figure out what to build next
Hey thanks for doing work on that. I think AI in games is interesting and I can’t wait to see where it goes. (Like, enemy AI, not ai being used to make games)
Perhaps one of these day, a game developer will finally decide: hey, let's make this actual Programmed Opponent a plug-in type mod so our fans can spend 12,000 hours making the computer routines more ... robust.
@@Sanvone well there is a difference between "I want enemy who actually acts like they know what they are doing" and "suddenly my enemy has been replaced by a love child of Hannibal Barca and Napoleon Bonaparte."
Arumba literally fixed it, and it started building, deving, spending its monarchpoints reasonably and even did some logic for wars and stuff. he submitted it to PDX, listing all the improvements that he did and yet they did nothing with it.
Reminds me of the time an 'Alien: Colonial Marines' modder changed the word "teather" to "tether" and it has significant positive impact on the game. It didn't magically fix the entire AI but the fact that a single typo fix made an impact on the AI, that's clearly noticeable, is kind of crazy. (Because the code for "tethering" the Xenomorph into a room now worked, they were able to flank and use cover instead of being completely broken.)
Also Paradox seems to have a long-standing tradition of ignoring their QA team. When they got axed, most people's reaction was "they had a QA team?" Technically this is still hearsay, but people attribute this revelation to DDRJake, who at first worked in the QA for EU4, and later was the game director for the title.
@@Xazamas i havent heard that name in a long ass time lmao. but from what i recall DDRJake got essentially all the critisism for everything even though the game direction, content wise, was pretty sound.
You made me imagine the horrors of forcing a 10 year old having to play eu4 to live, and do it badly too (they get exexuted if they do too well and piss off gamers)
@@arbikiuszelemelek3347imagine being forced to play as native America and do nothing. Or a 1state nation. This would create a new generation of traumatized children.
CK3 has the benefit of giving its leaders traits. When I play I don’t question or feel like the A.I is competent when the paranoid, lunatic, measles riddled bisexual 30 year old 2,3,5,1,0,0 king of sapmi raises it’s army of 400 levies on it’s capital as my viking hoard is approaching. I see it as ”lore accurate”
An interesting point is that Stellaris Leaders also have traits. But, has far as I'm aware, those traits only have mechanical effects, and don't alter the way the AI plays.
When it comes to the eu ai, I really like the idea of yeah make it smarter, but then nerf it through having personalities based on the ruler like in ck. when you know one ruler is very capable and aggressive be on your toes, but if they then die and leave the country in incompetent hands you get an easy time picking on them.
The AI's decisions being stupid can be justified in different ways depending on what you expect in that situation. A feudal ruler being a poor diplomat or a simply terrible strategist is realistic and expectable. Republics being commanded by rules who do not make optimal economic decisions is justifiable, it happens. The Problem therefore with stellaris Ai, is that there is no justification, no explanation that could possibly explain Empires of billions of sentients, with the ability to colonize the stars and extremely advanced Research AI to not understand how entertainment complexes function.
Maybe EU IV AI is bad, but at least it is not CIV VI AI. In that game you set the AI to the highest difficulty and when you get to the modern era or so you look at the strenght of that huge empire and they have a whoping 3 obsolete military units or so. You can literally conquer them with no problems with them being 1-2 technology eras ahead of you. The only time AI is difficult is at the start since they start with like 3x your units plus get extra cheats(to compensate for them being terrible I guess).
@@JakeBaldwin1I still play Spearhead for COH2 against Elite AI with friends from time to time. The amount of swearing and anger from 4 young men over the internet is honestly impressive.
also yeah i'm one of the people who play eu4 on very easy 😭i've played a few playthroughs with minor nations on normal/hard but atp i don't want to micro my wars too hard
It would be trivial to make an AI that can beat any player 100% of the time. Just make every non-player country no-cb the player on the 11th of december 1444.
here's the thing: most people don't actually want perfect god-like AI, they want a challenge that they can overcome. The goals of default strategy AI are to 1: to perform well, even with a game populated entirely by AI. 2: represent a challenge to any players. 3: be a competent ally, or have coherent diplomatic strategy so as to not frustrate the player (unless backstabbing is the point). 4: not be unrelentingly ruthlessly impossible to beat, leaving no opportunities for a player to actually play the game. And there's even a 5th one in historical games like eu4: relatively likely to actually do historical things, or do unicorn/easter eggs occasionally. This really isn't as simple as "make all 300 nations have perfect behavior."
To be fair, I would really love a "historical focus off" button in eu4. I agree that the AI should be more historical in general. I had several problems seeing Portugal AI colonizing Brazil, which is literally their main focus...Also, I have never seen GB go for India. But, as I already said, if we go more historical, I would also like an option that goes less historical. Just like Hoi4 has this button.
As a Hoi4 player i find the AI frustrating in different ways but mainly working with the plane/ship and tank designers as it will make the worst designs imaginable which means that with the right design you can beat it 1:20 in terms of production cost (when using planes as an example). This is dumb and i think i would feel a lot more challenged and have more fun if you could exploit the ai so easily
Depending on how much of processing time ai decisiinmaking of all hundreds of nations takes increasing its complexity might also seriously slow down the game
I think to improve performance there should be two gameplay levels, GS (macro) and RTS (micro). The RTS level is more efficient but takes more time and computing resources. Players and AIs directly involved (neighbours, vassals, enemies in war) with them should play on this level. The distant AIs should instead play on GS level. If I'm playing as Albania or Switzerland (or even Great Britain), a battle (or even a war) happening between Ming and Oirat should be something neither myself nor my computer should be concerned with. In such a case, instead of using RNG for every single battle there could be a single randomised calculation for the whole war. Instead of having the AI develop each province separately or build every single building, there could be a whole area development. We could even completely remove computations for distant AIs by having predefined behaviour for them. Distant in this case though would mean that they and the player are not affected by each other's actions in no way whatsoever. The switch from no computation to GS level computation could happen when you get close physically or when you interact with any of their neighbours. In the first versions of the game there were much less provinces, but this way there could be way more. It's just that the majority of them would be dormant for a huge part of the game. Something I think the EU series could take advantage from the CK series is using internal vassals. Yes, you have a much smaller demesne and less direct armies, but you would also use less time micromanaging and your computer could perform better, especially when you have a very big country.
Which is why mods like Starnet are absolutely essential when playing Stellaris. Unless you want to fight authoritarians because Starnet gives full citizenship to everyone regardless of ethics so it just feels like you’re fighting the evil oppressors but they treat their pops the exact same except their leaders wear shiny hats.
Of course it gives full citizenship - it's always the best option in the long run. There is no benefit to maintaining any kind of underclass, as all it does is make your downtrodden pops less productive. Which perhaps reflects reality, but it ignores the real-world political pressure of social elites, to maintain the status quo even if reform would be a net good for all. Stellaris is full of these kinds of non-decisions, where there is no reason to take a short-term benefit at the cost of the long-term. Because Stellaris, like many other strategy games, is a progress race. The empire that grows fastest will inevitably dominate all of its peers. So the player is pushed to always prioritize long-term gains, to keep up their place in the progress race. Keeping the commoners in their place might make sense to a short-sighted elitist, but to a far-sighted player it's just a massive opportunity cost. Crusader Kings avoids some of this, by keeping the player rooted in the present moment. The player has far less reason to focus on long-term goals, because the simulation is so chaotic that there's no guarantee that any long-term investments will pay off. So the player has to just focus on their current character, and play the game life-by-life, adapting as they go. It's not perfect - the player's ability to Quantum Leap into their children produces dynasties with impossibly consistent long-term goals. But it's still vastly better at producing a historical narrative born of short-term decisions, than other strategy games.
Except that Starnet hasn't really been update to support new game mechanics/changes for years since the original mod author is from Ukraine and can't update the mod (it's only update to run on newer version but the AI hasn't had any change for a long time).
So my paranoia of the multiple moments i sent my armies to stop a siege from inside, what would be, the AI fog of war for then the AI stop the siege as soon as i unpause and proceed to manouver all around the entire mediterranean sea was not the AI cheating but the AI being successfully dumb enough to bait my army and then plant the seeds of confusion on me about their army location. 10/10 AI
The "AI cheats with fort zone of control" -cope probably stems from the fact that the ZoC rules themselves are at least a bit convoluted, but provinces have very irregular shapes and formations that can further lead to confusion. Worst offender is one province in Burma, that is so small and sits so snugly next to the Ava's capital level 3 fort, it literally is inaccessible until the said fort is sieged down. This effectively grants the AI a safespot where to rebuild armies in. I'd argue that the patch where AI went way overboard in deleting forts made the 'The First Toungoo Empire' -achievement easier by an order of the magnitude.
I just want 5-10 AI personalities instead of three levels of aggression. There is absolutely nothing wrong with a wildcard AI either. Civ Gandhi's kindness overflowing into 11 out of 10 aggression is a great example of a community appreciating such.
7:38 the ai in ck3 is fine until a crusade happens in which it will become insanely brain dead and will play like rodents running away from winning battles and baiting you like rodents.
Crusades are mind numbingly stupid in ck3. We had all of Israel captured, and what does the AI do? Surrender! Like wtf we had them on the ropes. Tons of gold on the treasury and we walk away with nothing.
Looking at how they develop their provinces, yes. The way I develop them, is specialise them either into money-provinces (good trade goods : workshop manufactory trade and mainly dev production) and manpower-provinces (weaker trade goods : barrack training fields... and mainly dev manpower). Meanwhile the AI doesn't seem to focus their provinces. This means that at equal dev and money spend, my provinces are overall more effective as everything synergises nicely, while for the AI it all looks random and unfocused.
your strategy only "works" precisely because the AI is bad. If it were competent, the population of the military provinces would decrease over the years, as everyone would migrate to the developed provinces, since no one wants to be condemned to spend the rest of their life in a miserable place until they die in a war. In a few years, you simply wouldn't have anyone left to recruit into your army.
Somewhat related: One of the many things holding back the 1999 'Sid Meier’s Alpha Centauri' is that the AI isn't that good. (More imminent one is the archaic UI and control scheme.) However, someone raised a point about that this isn't because 1999 technology didn't allow good AI - the AI in more recent Civilization games isn't any more advanced. They've just started designing the games around the fact that AI is bad at tasks like long-term planning and managing very limited resources. Aside the main series never having some of the SMAC features (unit designer, orbital infrastructure, dynamic terrain/weather) the following were removed after Civ III (that was developed somewhat concurrently with SMAC which used heavily modified Civ II engine.) Civilization IV removed the city-based supply (have to pay shields/minerals for unit upkeep every turn) in favor of global unit pool with certain number of "free units" and rest are paid with gold. Wealthy-Science-Luxury-allocation was also removed in favour of a simple slider on how much (0-100%) of your gold income you invest into science. Civilization V completely divorced Science and gold production. Units now always have their per turn gold maintenance cost. To bring this comment back to the Paradox games, it could be that Victoria 3 trying to remove the traditional army stacks was at least partially motivated by this principle. AI playing the game "wrong" makes for an extremely frustrating experience (either your ally is useless or the enemy army marches across the entire Eurasia to either run away from you or siege a remote region.)
I wouldn’t really say Vic3 was made with improving AI in mind, it still screws up in wars and mismanages economy to the point they get easily eclipsed by the player sometimes more than a hundred fold
@@newguy8288Depends on the situation though. I actually have a harder time against the AI in Vic 3 than most other Paradox games simple because they love to join the opposite side of a war.
I recall playing a strategy game from way back called "Heirs to the Throne" and it was notorious for having an unforgivingly strong AI. There were no settings to adjust the AI strength, but the rules of the game could be changed at creation. The AI however never made a mistake and very carefully read its odds and even under favourable starting conditions, was immensely challenging to fight, and this is with only one active. The game supported up to three AI opponents as well as the unaligned forces.
In case of Stellaris and HoI4 the death of AI was easy to observe as both games had evolved too far from original framework. Especially in case of Stellaris. It's like trying to make Counter Strike AI play Dota 2 just because they're on Source. And I can't suggest a counter to this as evolution of gameplay was in general a good thing.
The issue was the shift from a tile system to the convoluted district/building system we have now, which has two parallel systems doing essentially the same thing at the same time.
As a total war player I just want CA to make a system were the AI has a rock,paper scissor system that says to the AI that charging all your cavalry into a pike wall might not be a good idea or maybe force the AI to actually use formations and make the formations used unique to certain factions or change them based on what the AI army has and what the player has and the formation the player is using because why is 40 units just charging my spear unit and allowing my archer tower to get 300+ kills
Add just STANDING STILL while being erased from existence by artillery. Problem is not that they let us win - problem is how stupidly easy it is to abuse them instead of outplaying them.
AI need to learn counters in like every game because for some reason a lot of AI is made more difficult to beat by just giving them insanely dumb buffs even in all the total war games, and it really should just become smarter, AI would be so much harder if it did know to counter cavalry with pikes and pikes with archers and everything else instead of it just getting +1 billion money and 15 armies of elite troops with +50 attack and defense.
I will say, the god awful AI of HOI 4 has been mostly fixed with the Gotterdammerung dlc. While certain aspects are completely broken (fleet submarines evaporate everything. There's basically no reason to build any other ship once you have the tech), the army ai has improved significantly and actually moves armies by concentrating forces at key points to attack. Makes for more dynamic gameplay and actually challenging at points, even as a major power. I wish the different difficulty modes affected ai behavior though, instead of just giving ai buffs
Nah, the hoi4 ai is still brainless, just a tiny bit less now It still cannot design even half decent designs for tanks,planes,navy or templates, it also cannot really use MIO's properly
Well from the game developer pov, "smartness" of the AI is only one of the concerns. The most important one, being performance of hundreds of individual AI actors making their calculations in real time. Because some advance stuff like learning models, require a lot of power to run even one of them. The second concern is, that's you can't make AI, to independent. It's seems counterintuitive, but complicated and independent AI algorithm without a clear boundaries is a pain to debug and fix or even change. What if AI will find that's the best war winning strategy is split and merge it's troops at the speed of frame rate or even delta timez which gonna make battle eternal? Or even just crash the game? Well and the third concerns, is that's from the gamedev pov, AI goal is not to win, or be smart, but to entertain the player and provide illusion of smartness or challenge. Some objectively better strategies, just isn't fun for the player.
One thing that always bothers me in eu4 is the ai being almost completely immune to land attrition. I’ve gone out of my way to avoid going over supply limits, stacking attrition loss reductions, etc. And I still take 10x-100x the amount of losses from attrition that the ai does. Meanwhile they’re siege a fort in Siberia with a 100+ size army or marching doom stacks everywhere.
Thats because AI can't handle attrition. Paradox very early on decided that AI is able to cheat at some things, otherwise it just breaks. Military Access for example. The AI doesn't use diplo slots for military access. Why? Because it can't handle it properly and will destroy itself with diplo points or not be able to wage war 90% of the time. (its also good for the player, since you can move with their military access and dont need to use diplo slots for it)
@OlafJorigson It felt appropriate to mention since it was brought up at 20:48 they don’t take naval attrition but didn’t mention land attrition. It may be unlikely to lose a war because the ai ships didn’t take an extra 5% damage at sea, but I can easily see a war being lost because I’m taking an extra 50%-100% manpower due to attrition just from sieging castles. I once had a war between me + 2 ai vs 2 ai. I let the 4 ai fight each other while I just captured castles and avoided any fights. I still took just as much casualties from siege attrition that each ai lost (individually) to battles.
I would argue that the arguement at 5:30 is wrong. I haven't played rocket league but with convolution and an autoencoders it shouldn't be hard to compress a complex 3d environment down to a vector which can then be used as input to an NN. Action games environments aren't magnitudes more complex than grand strategy game environments.
Yes but they are infinitely more random and difficult to predict. Whereas in PDX games everything can be solved with a simple weight on a decision An easier example is hoi4: there are mods that just add a bunch of triggers for the AI thanks to which it becomes almost as good as an experienced player, and even better than Xorme when it comes to differences between vanilla and mod. You get to fight blursed builds like 100 div mass mob Poland, ~15 heavy TD divs france, or a meatwall of 10mil red army troops in 41. All that with a few megabytes of code (probably even less) Now try getting anywhere with a wall of "if"s in Rocket League without it taking up 200 yottabites of storage. It's much more difficult to make a good action game AI, and having to use a NN as you said is probably proof of that
@@maciekGTR I don't think you know much about machine learning... You can look up function approximation methods for reinforcement learning. From what I have observed, rocket league environment can really be described as a finite list of elements in a finitely sized 4D space, where each element only has finite parameters. You can easily compress this down to maybe a few billions states. It's gonna use a few PB space at most. And then rocket league has a smaller action space then EU4 too.
Shame not to cover the AI in HOI4. Definitely a contentious one. The recent force concentration changes are amazing but it still leaves a lot to be desired, and frankly in HOI4 I'm 90% sure the AI can cheat.
I wish you could stack TCA that high. But you can't even really stack it, because there is only one modifier of 40% TCA in national ideas. 😢 And only for RNW nations or custom nations.
love the measured, fast narration and focused, sober visual aids. too many people do "streamer voice" or make "funny exaggerations" every other sentence, fill the screen with wacky memes etc
Major powers breathing down your neck isnt just a roleplay element of playing as the Dutch. It is the defining characteristic of the Dutch in this time period. You thought we went all the way around the world for fun? Anyway, I think there is a lot of improvement to be made in games using neural networks for the AI, just to make it play more like a human player. Train the easy AI by having it play against Bob the intern with 100 hours in map games and train the expert AI against people who have nolifed map games for 10K+ hours.
No indeed, but I said it as moreso that in especially older Eu4 versions, the Dutch land was so good and the AI was so incompetent, well, you became the Major really quickly
If you don't want anyone actually playing the game that's what you'd do, because a neural network AI would exploit the shit out of every mechanic in ways that players can't.
@@Gustav_Kuriga Well, considering its a game you can pause at any time and is semi-real time, semi-turn based, there isnt anything the AI can exploit that people cant. You could theoretically pause the game every tick and give micro commands to every individual army unit. Of course they AI will have to get some limitations on how much effort it puts into the game. But the AI that is only trained to beat Bob the intern wont need to exploit much to do so. The expert AI might learn some exploits, if that becomes an issue you can just forbid it from doing those actions. But players use exploits too, so i dont really see your problem here tbh
@TheSuperappelflap You do realize games can have bugs/exploits that only appear when you input commands at thousands of times per second right? Do I really need to explain that to you?
An interesting thing with Paradox sieges in particular is how often the mechanic gets re-used. The mechanic is "Roll 1d14. Add modifiers. If the value is greater than X, complete the siege. If the value is greater than Y, add +2 to the stacking modifier value. If the value is greater than Z, add +1 to the stacking modifier value." But that EXACT system gets re-used all over the place. It's the system used for First Contact and Archeology in Stellaris. And they ripped it out of the Europa Universalis board game. EU4 is just the first time they revealed what was actually going on under the hood. At least that was according to one of the people in charge early on in EU4's life cycle (pretty sure that was before the Conclave DLC for CK2, but I might be wrong). Given that mechanic, it probably shouldn't be too hard to train an AI how to evaluate sieges. Especially if you're factoring in "Is it worth losing the modifier I have in order to go do this other thing, and potentially come back with a bigger modifier in this other area?" Though you would also need to factor in time and such. And you could probably use that for the AI to evaluate whether or not a siege is actually worth attempting with the forces it has available. I don't know the current modifiers used in EU4, but I suspect that training problem is much simpler to approach than the full game. And having a neural network work out a system that would work for those evaluations, with minimal changes to the codebase, would probably lead to some helpful patches. I don't know about the rest, though.
I think another issue with the Stellairs A.I. is that over the years more and more stuff was added to the game that rewards planning ahead and understanding how things work together. Things that a player can effectively use very well but a A.I. struggles to even begin to under stand. The biggest sign of this defeat is the "difficulty scaling" option, which really is just a band-aid on a heavily bleeding wound.
Machine learning algorithms are incredibly computationally expensive to ship into everyone’s games, yes sure you can train it beforehand and ship a set of weights to the players instance but Therein lies the major issue. The ai requires hundreds of unique “thinking equations” which determine the mathematical value of each potential action, then it must choose the highest valued choice. Eu4 has hundreds if not thousands of game state variables so you can imagine it’s hard to formulate these equations and add weighting on top of them in a balanced manner. The equations also have to be consistent between themselves ie if the think equation for building a granary naturally increases faster than the one for building a new unit it will be skewed. This is from my own personal game dev experience, I cannot imagine how hard it must be to make these for eu4. A challenge to you lemon, try to formulate an equation for each potential ai action to gain a perspective on the developmental difficulty of this
One of the problems is the lack of communication between the player and the bot. I've had this issue as a hoi4 player, where the AI covers fronts I've already taken care of but doesn't care about the ones I cannot spare soldiers to defend.
Im glad someone pointed out that it's not really AI and it's all just machine learning. I can't tell you how annoying it is when you hear all the media, Nvidia, CEOs, and youtubers going around talking about "AI".
I never really saw that AI fort problem in EU4, the path finding and making decent fort lines that actually lock down an area is pretty easy to understand for me.
it's basically "it's bad to make player feel bad at game" combined with some performance considerations. Developers could build AI that is at competetive level or near it - especially with good micro. Just look at AoE2 HD edition hardest AI. It's quite strong on competetive level and this game is also very complex technically
I had CK3 AI within life of a single character conquer almost the entirety of europe except Iberia and british isles. Some random swedish guy got Great Conqueror and Scourge of the Gods.
I always argue that the issue with games, EU4 especially, isn't that the ai is bad but that the games have been built up to be too easy or too boring internally. It frustrates me so much that EU4 ai is sometimes megablobing, like getting the russian 1910 borders in 1600, because I then have to blob hard to keep up. I want nothing more than getting the french borders in 1789 to be something that is both hard and satisfactory to have done. Its so annoying that Worlds conquest by 1600 is commonplace.
The current AI is good enought for people who play on very easy or easy. But for people who want to face someone with some competence at the game, it just fail to do so, it is almost as it is programmed by someone who isn't particularly good at the game and just wants the AI to do stuff that doesn't look like a bad decision at first glance without the context of the situation.
I had a run on shogun 2 as the Takeda. Was very early in the game (less than 20 turns in). An enemy clan marches a full stack and guns it for my territory. I was forced to basically reload just to prepare to fight this one army by scraping every man I could and forcing a battle which I could gain every advantage I could. I was against some really good troops and lost once or twice. I had to reload as if I lost, my campaign was over. I was able to finally win a pyrrhic victory and the ai army retreated. Although I had to save scum, it was EXHILARATING to experience and overcome the challenge. Now, would I want that every 10 turns? No! The ai has to be Stupid sometimes so the player doesn't get absolutely stomped. It should also be smart sometimes. I think its hard to make a scenario that lets the ai be smart without removing lower skilled players from being able to enjoy the game
Shogun 2 AI was good in that it simply was aggresive. Did well to consolidate late game (lowering endgame turn times), before Fall of Samurai had workable naval AI (that simply attacked thus putting player instantly in pressured situation) and was kindly offering player epic castle defences (meat grinders) that were still something you had to work through but were satysfying :).
That deck of cards thing would ultimately push some crazy players into waiting the bad thing they don't want to be played before they engage in important battles :')
To add Vic 2 to the mix, they AI in that game is painfully awful at waging war. They will randomly seige provinces early game when the first priority should be army concentration destroying enemy armies, and late game it is so easy to lure it into battles. Like, you put a small army on a mountain tile, the AI will attack it, then you reinforce it and win, and when you attack, the AI won't have a backline to reinforce. The AI also leaves provinces open late game even if they have a large enough army to create a frontline. Outside of fighting wars, the country AI seems fine, but man, do capitalists suck at being capitalists. They lack the ability to optimise production like in Stellaris
In Stellaris the AI sucks because it doesn't know I am minmaxxing with 6 word documents everysingle day that passes. Also they don't plan ahead, only if there's a sudden unemployed pop or a deficit of some sort it starts to do something about it. Otherwise is like: "I sleep."
For the dice roll randomness, the Fire Emblem series has had a number of neat math tricks to make it closer to how humans think. The most basic from the GBA games was just you random roll twice then average the results for the true number. With how it uses the rolls, this makes high percentages (90+) nearly always hit as the average is below it, and low percentages (5%) happen extremely rarely. Some of the more recent games have this formula become a bit more complicated, but it still shows how you can make randomness work closer to how humans expect it without doing the "Karmic Dice" approach which just end roll streaks.
That's very fair, and indeed humans forget that a 5% chance happens once every 20 times: so 1 finger represents 5% of their fingers and toes added up (helps with visualization I find)
While I agree with a lot of what you said, I think your argument against why AI can't be taught how to be good at grand strategy games is missing a major element - the scheming or plotting or even just planning. You could also call this 'the human element' or just strategy, but AI struggles with both judging their opponent and thinking ahead. You touched on it briefly with the player knowing that another player would have 300k more soldiers hidden, but didn't really elaborate. Ultimately, if they made an AI that was exactly at my skill level, I would still win. Not because I'd just be better than the AI, but because I can lie, trick, bluff, deceive etc. an AI in a way I cannot do to a player and an AI cannot do to me. Very few AIs for video games have long reaching planning capabilities, instead being mostly reactive to the current scenario, and fewer still have a memory or the appropriate reactions to use it. For a stupidly basic example, in HOI4 if you ever find yourself up against the AI in an unbeatable position, often your best choice is retreat. Not because you've been beaten - but because if I fall back, the AI will see "oh, open tiles!" and walk into them, claiming them. It will move its entire force with the frontline and leave no-one guarding the unbeatable position. Then, I counter-attack quickly, and while his divisions are low on org and retreating I attack the unbeatable position again, now guarded by the broken remnants of his retreating force. A player would (almost) never fall for this, but the AI would. And sure, you could have the AI not advance into suddenly empty tiles - but what if I ACTUALLY retreated because there was an invasion on the other side of my country, and advancing was the correct move? The AI still has no idea where my units are, but a player could put 2 and 2 together in a way the AI can't. A player knows someone who was just winning suddenly abandoning their frontline and disappearing is a suspicious act, an AI just sees it as no longer knowing where those units are. It's incredibly difficult to code an AI that can handle something that a player would do naturally, and I would say that there's no AI in these sorts of games I can think of that is immune to these tricks. That's why I find Chess a poor argument for why AI for PDX games is hard. Chess is easy for a computer to win, it's very rigid in terms of what each player can do on their turn, each turn is in order and each turn changes the possible outcomes of the next turn. I code AI for my mods for ArmA myself and yeah, it really is just weights and decision trees all the way down, but a lot of my design work is around making the AI behave the way a player would, rather than trying to make it outright "good", and so much of it has to be faked with subtle cheats or forced behavior of the AI in order to achieve what looks like 'normal gameplay' to the players.
For stellaris in particular, I think it's that the AI gets more short term benefit from stuff like strategic resources so it goes hard in on that and ignores everything else.
3:54 If something can show that they can make a good AI (as CA showed us a decade ago) is funnily enough a different thing with the initials AI Making a good AI for a horror game is very much a harder challenge that doing so for a “spreed sheet simulator”
I've long been curious how machine learning AI would play games like this or Dominions. Would be a lot to learn from such a thing. Good luck getting the resources to train one sufficiently and update that with patches, though.
My biggest Problem with most grand strategy game AI's is that they don't scale with the player. Individualy, but also over the course of the game. Which leads to hard early games and easy late games in nearly all cases. Total War is a great example of this. Early game on higher difficulties is really hard, because AIs on the other side of the map will drop whatever they are doing and send dedicated death squads to the player wihtout ryhme or reason. While late game is really easy, because the AI does not dynamically build up anything to stop your snowball. In short, the AI does not adapt its actions to the strength or playstyle of the player. Which feels like it should be really easy to fix: * roll one or multiple random rivals and cheat for them to grow at a similar rate as the player. * adapt these rivals to your economic and/or military rating
18:46 I do, just cause I hate the random 9-9-9 in important battles. Like, that is just completely random and if it's a big battle, tough luck, nothing you can do, which can be frustrating instead of being able to see: "Hm I went in and they had reinforcements waiting, that I couldn't see. I should play more defensive."
I recall there being an issue when zone of control was introduced (eu4) where if the AI could find a land route to a province, it could move to that province and ignore zone of control. I definitely witnessed this issue watching the ai walk right through barrois to approach paris (metz -> barrois -> champagne) instead of going the long way through belgium that any player would have had to do. I would have confidently said that the ai still does this behavior, but thinking about it I haven't seen them do it and i've been playing a lot of 1.35.6 recently. I wonder when they changed that. Meanwhile the ai absolutely knows exactly where my drilling armies are at all times because the moment I click drill because I don't care about a war i've been dragged into a 30 stack of random germans manifests on a fort-less border with their infinite military access to beeline at my 0 morale army.
On the deep sea atrittion thing I could see a situation where you have your army on boats run into an enemy fleet slightly weaker than your fleet but you still don't wanna loose ships fighting so you kite the ai fleet ending up in deep water and after a few turns they catch you and you lose the fleet and a nessicary chunk of your army snowballing into losing the campaign
The fundamental problem is that the initial statement of devs being just not good enough is true most game developers dont understand their game even to the point where something that is clearly op or up does not get balanced for literal years and dozens of updates. Honestly the easiest thing they could just do is make the interface that allows people to code and or train their own ais.
I thought Stellaris’s “AI” worked pretty well with the old tile system and admin caps. It would still do wonky things but you couldn’t outcompete it anywhere near what you can now. The amount of min maxing you can do is just insane. An interesting factor in “AI” is how many opportunities and options does a human have to outcompete it?
19 днів тому
The thing I hate about AI’s in games, is as soon as you ally them they become brain dead. Its like they literally break, or they’ll never ally with you despite being surrounded by 1 million troops of mine and only having their king left alive. Also hate when they run their troops around and never engage in battle, they just back cap you and flee etc. They don’t have to be as good as players, they just have to “roleplay” the nation they’re taking over. With some variability so its not always the same build every time.
It’s funny but I find the AI is EU4 is pretty competent it’s just that the players threshold for doing everything well is just much higher. On VH a big austria is a massive pain. Anytime you leave anything short of an 80k stack somewhere they will come and hunt down your armies.
A couple things with the AI in Stellaris. First of all the sector management has been fixed, one of the basic settings (that should always be turned on) with the management is to build according to planetary designation. If you set the planet designation to mining world the AI will build mining stuff. Second of all in slight defence of the AI not specializing it's worlds. A player is very capable of adapting to circumstances and fast, but say an AI loses a primary mining world it's economy would be utterly crippled and it'd have deep difficulties fixing that if you tell it to specialize it's worlds.
Second of all in slight defence of the AI not specializing it's worlds. A player is very capable of adapting to circumstances and fast, but say an AI loses a primary mining world it's economy would be utterly crippled and it'd have deep difficulties fixing that if you tell it to specialize it's worlds. Especially since the AI is incapable of demolishing anything.
The AI can be brainless, however in an older update of Hoi4 there must’ve been hidden stats because I played exactly like the AI in historical for 2 nations, yet lost both times against each other.
Literally from the 1st statement. How can you make it if you play they game. 1) In general, AI obeys zone of control rules but it also cheats them sometimes in a way that players can't. 2) devs can't create a CPU player that would beat 99% of players. For them to be able to do that they must be in 99 percentile of player base (which they aren't, they are casuals). Even with dedicated machine learning Alpha Go and Dota AI that were trained on millions of matches were then beaten by average players when strategies were established. 3) In CK AI indeed behaves more on its traits than surroundings. And it's really good. My personal gripes with EU4 AI: 1) It tries to contain player from day 1. 2) It can see though FOW past 2 provinces. 3) Diplomacy. 19:18 that's called triangular distribution)) Also they've implemented that system of not repeating rolls in Baldur's gate 3. 21:18 with that I'm going to argue. Imagine how you get FOW in enemy territory. For 2 layers of provinces so does AI but after that it can see everything. Put a doom stack deep in your territory and move it around. AI would react to it and move its armies away. To make it obvious use speed buffs or cheats for instant movement.
1. Again... no it doesn't: I promise you. It doesn't. 2. The people that made deep blue where not 'smarter' then deep blue at chess: what kind of logic is this?? 3. Here we agree 1. I don't really agree not going to lie, but its another (yet again) optics problem 2. It does not. Watch the section chief. 3. Yeah honestly no clue what it does there
@@LemonCake101 I am sure AI can see the armies because if i click on a province to engage enemy all the way from siberia, the enemy stack moves, despite being stationary before
@@LemonCake101 Not that I agree with the OP, but EU4 "AI" is just a bunch of if/else statements whereas Deep Blue is actual AI that recursively self-improves. They have virtually nothing in common.
@@LemonCake101 well, chess have an simplistic rule set. Where AI has extreme advantage over low "human RAM". Plus game developers (as you've pointed out) don't create an AI but a bot. In games where rules are constantly tweaked from game to game human player has an advantage over AI. With player containment I got an easy example: when you start as a small tag (can't rival majors) majors would ally your neighbouring OPMs, but as OPM you can't ally those majors 1st month. On FOW, when you get time, try to set up an experiment (establish delineations before start).
Complexity of how many mechanics you stack into a game for sure can hurt a.i. in how it creates allot of moments for cheese to happen from interacting with anything they not patch enough or does fly under the radar. The problem with a.i in videogames is usual that it does not get updated or maintainted very well and players learn its weakness.
Yes, this has always bugged me! Seeing how well Alphastar does at Starcraft makes me wonder how incredible it would be if Firaxis unleashed something similar on the Civ franchise. It could be a big marketing gimmick too!
24:09 Actually planetary automation does not work anything like the AI, instead the only thing planetary automation does is build distics/buildings based on your designation whenever there are unemployed so on a mining world planetary automation only ever builds mining distirics or the building that impoves mining, that is as long as you turn off the option to checkbox that let's planetary automation build strategic resource buildings wherever it wants to build them. All you need to do is 1. Manually set your designation 2. Click the settings wheel icon on the planet automation thing and set the settings appropriately.
A potential technical issue isgame performance. The ai is already one of the things slowling most pdx games down the most, so it being light on performance is one of the reasons it is so simplistic.
Right? I am so sad about this, recently I upgraded to win10 from win7 and when I went to play it crashed like every 25-40 minutes? Such a good game, Such a shame really.
You don't mention most important thing: performance, it's possible to make good ai, but probably it will make games unplayable even on NASA pc. And CK ai win at this category too because it doesn't struggle with late game performance.
Performance does matter, but you can make a really bad AI with bad performance, and a good AI with good performance too. They may be connected, but not 'inherently'
I absolutely agree with first part, but don't think its possible to build a good ai with high performance, especially if you rely on ANN, it's just math.
What I must say about AI and the rocket league thing is that AI is naturally gonna be balling when it has all the information like in chess and rocket league, when it doesn’t have all the information its a different ballpark.
Victoria 3 AI is pretty decent these days- diplomacy values are pretty straight forward and react to things that you would think they should react to and does its level best to explain them. AI is annoyingly good a military even if it doesn’t use meta templates and it can build in the economy well but not total meta
The only "AI Cheat" I feel is the lack of clarity in troop/ship limits or what not. For example, if you reduce Venice to only two provinces, they are still able to maintain a fairly large naval force limit. Secondly, I can feel confident infighting a nation that has 40k troops and 20k manpower, then declare war on them. However, theres always some variation in the amount of extra troops they can raise, and it can feel absurd when the AI has 40k troops, 20k in reserve but the next month their troop count jumps and with three mercs stacks to be at 72k. Like were they not near their force limit at all or is the AI that willing to go into debt?
There was a recent interview with a former Creative Assembly AI developer who claimed that he was told not to make AI too competent so players wouldn’t have a challenge, though take that with a grain of salt. I think best example of a strategy game with a competent AI is GalCiv 2.
Thanks to Factor75 for sponsoring today's video. Use my link to get 50% off and free shipping on your first Factor box! strms.net/factor75_lemon_cake
Just make you own food ? İt's not even good to begin with from looking at the reviews atleast it's not fume or some other thing that literally worse that it's counterparts and also much more unhealthy (plus it's very expensive)
@@Just_Nice07 sponsors aside, ofc making your own food is going to be better, but this is going to beat most takeout price and quality wise, and much like most things in life the premium is for the convenience
@@LemonCake101 👍
Is not ai art .But ai image/video imagination generator .
Because it only generator imaginary From the person who Enter the prompt
When I beat the AI it’s completely brainless, when it beats me it’s cheating and broken. Simple really
Well I'm human so I must be right
" It is brainless if more people than me reliably can beat it and I'm no longer the best strategy gamer around :< "
I mean, those 2 statements are technically not contradictory.
@@ls2260 When it comes to micromanagement, AI can pull of "quantity has a quality of it's own." By sending a small stack to burn down player's countryside, the amount of attention and effort requires to take it down (or at least chase it away) can feel disproportionate, if you also have to manage to actual/proper warfront as well.
Of course this is probably a real strategy that has been used to distract/exhaust the enemy. Difference is that Ottomans never sent an army to do that to the Russian Far East.
Based
The thing that makes CK's AI 'good' is that usually, when I am questioning a country's decisions or impressed by how well it's doing, I just look at the leader's traits and go 'Ohhhh, everything makes sense now'.
Childless Catholic ruler not ransoming his wife because he is greedy lol
It’s also funny how absurd it can be. You can challenge a crippled old man with a prowess of -2 and he’ll accept because you killed his son last July. Despite the fact that you Will absolutely kill him.
Likewise, you will invite shy characters to a party and they’ll say no.
@@mahmud7645not ransoming his wife because he just got a ticket to a free divorce 😎
@@xtxownagetbf if you killed my son im throwin hands wether youre gonna obliterate me or not
Also, in history... rulers could be idiots.
Allied AI: Runs around in circles with 10 seperate 4k stacks that get stack whiped
Enemy AI: Marches around in 40k hivemind death stacks that cross siberia to siege down a human player
Paradox: AI doesn't focus the player.
AI: Random alliance that exists only to screw with player.
Or when you're playing as an indigenous tribe on the West Coast of the US, and suddenly Portugal is coming in 1528 and they already own all of New England and Rio de Prata.
@@domehammer France alying Venice and Genoa and Austria allying Milan - just beacuse I decided to play Florence...
One time when I was playing as Poland, I called Great Britain into a war. They combined their armies into a 120k stack, moved it to London, and spent the entire war dying to attrition.
@@RDA000 630
Everytime I take over an AI planet in Stellaris I *IMMEDIATELY* think of the
Joe Pesci: "What the fuck is this piece of shit?"
Factory world with 1 city district, no industrial districts, 4 agri districts, 3 commercial zones, 2 administrative offices and an alloy foundry
@@hogndog2339 sounds like my average planet
Ah yes good old CK2 Byzantium.
"Basileos, a 2nd revolt has hit the Theodosian walls"
Sept. 911 AD
Sept. 911 AD? I see what you did there.
September 11, 1001 AD.
Classic Stellaris experience: Conquer an AI empire. Spend the next 2 hours paused because you're essentially redesigning every single planet they have.
Stellaris made me a pacifist
Conquer ai empire: crippling economic crisis, everything besides food and alloys is now in deficit
You either go pacifist or go exterminatus
Literally, conquered my neighbour and I got an immediate -1000k in energy credits
@veryangryduckpl2122 Feudalism In SPAAACE! the AI get to keep their bonuses, and I get to rule the galaxy.
For context i make or help with stellaris mod making been at it for +3 years(played since launch). Especially AI integration. I felt obligated to explain why the Stellaris AI does what it does.
In short, the problems you mention are coming from one area which is there 'Economy plan'. This is what governs how big a budget is for what, when and where it does what. The catch is, a lot of that is based on if resource below/above [number], then give weight increase to secondary resources(Alloys, Consumer goods, Research or military ships). This means it can only make stuff when it needs or has a surplus of said item. It is by design incapable of planning ahead, only ever reacting to its own bottom line. Which means it can be VERY responsive, but has bad specialization as a result.
This combined with the fact that the AI is incapable of planning in correlation with its Research, makes specialization always worse when compared to a player. The funny thing about the Ai is, when you get the debug open and see what it wants. It makes sense what it wants. It is not wrong, it just cannot plan. Which makes it all the more tragic in my eyes. Like it knows it needs more minerals, but it cannot remember to have a dedicated mineral world. So it just dumps it somewhere with a research building because there was an unemployed pop on that planet.
Sounds about right... thanks for the insight!
actually this is how i was playing back in time xd
Apparently an older version of the AI built of sorts of planet templates that would guide them towards specialization. A bit like people (or at least I) do. The problem was that the templating was too rigid, and the AI would just conk out of it couldn’t figure out what to build next
Hey thanks for doing work on that. I think AI in games is interesting and I can’t wait to see where it goes. (Like, enemy AI, not ai being used to make games)
@@52flyingbicyclesCorrect, Plus it was separate from the budget of an empire. So it gave AI massive deficits if not over compensated for with cheats.
Perhaps one of these day, a game developer will finally decide: hey, let's make this actual Programmed Opponent a plug-in type mod so our fans can spend 12,000 hours making the computer routines more ... robust.
One day :)
Espiocracy is doing just that
Creative Assembly allowed in Shogun2 for every battle in single player to be joined by human as opponent if player wanted. Guess how popular it was ;)
@@Sanvone well there is a difference between "I want enemy who actually acts like they know what they are doing" and "suddenly my enemy has been replaced by a love child of Hannibal Barca and Napoleon Bonaparte."
@@Sanvone Not very, because connection issues? I'm guessing here.
Arumba literally fixed it, and it started building, deving, spending its monarchpoints reasonably and even did some logic for wars and stuff. he submitted it to PDX, listing all the improvements that he did and yet they did nothing with it.
Yeah, quite a few mods do too, but people don't exactly universally use them
Reminds me of the time an 'Alien: Colonial Marines' modder changed the word "teather" to "tether" and it has significant positive impact on the game. It didn't magically fix the entire AI but the fact that a single typo fix made an impact on the AI, that's clearly noticeable, is kind of crazy.
(Because the code for "tethering" the Xenomorph into a room now worked, they were able to flank and use cover instead of being completely broken.)
Also Paradox seems to have a long-standing tradition of ignoring their QA team. When they got axed, most people's reaction was "they had a QA team?"
Technically this is still hearsay, but people attribute this revelation to DDRJake, who at first worked in the QA for EU4, and later was the game director for the title.
@LemonCake101 do you have any reccomendations? I have ~ 3k hours
@@Xazamas i havent heard that name in a long ass time lmao. but from what i recall DDRJake got essentially all the critisism for everything even though the game direction, content wise, was pretty sound.
replace all the ai with chinese sweatshop workers
…
wait why did the AI get worse
@@LemonCake101You underestimated the capability of the slaves- I mean... the unpaid indentured servants of the glorious Middle Kingdom 😤😤
You made me imagine the horrors of forcing a 10 year old having to play eu4 to live, and do it badly too (they get exexuted if they do too well and piss off gamers)
@@arbikiuszelemelek3347imagine being forced to play as native America and do nothing.
Or a 1state nation.
This would create a new generation of traumatized children.
It's probably not worse than most office jobs. Hell, sign me up!
CK3 has the benefit of giving its leaders traits. When I play I don’t question or feel like the A.I is competent when the paranoid, lunatic, measles riddled bisexual 30 year old 2,3,5,1,0,0 king of sapmi raises it’s army of 400 levies on it’s capital as my viking hoard is approaching. I see it as ”lore accurate”
An interesting point is that Stellaris Leaders also have traits. But, has far as I'm aware, those traits only have mechanical effects, and don't alter the way the AI plays.
When it comes to the eu ai, I really like the idea of yeah make it smarter, but then nerf it through having personalities based on the ruler like in ck. when you know one ruler is very capable and aggressive be on your toes, but if they then die and leave the country in incompetent hands you get an easy time picking on them.
The AI's decisions being stupid can be justified in different ways depending on what you expect in that situation.
A feudal ruler being a poor diplomat or a simply terrible strategist is realistic and expectable.
Republics being commanded by rules who do not make optimal economic decisions is justifiable, it happens.
The Problem therefore with stellaris Ai, is that there is no justification, no explanation that could possibly explain Empires of billions of sentients, with the ability to colonize the stars and extremely advanced Research AI to not understand how entertainment complexes function.
It certainly doesn't help indeed, from an optics point of view
because it's not literally me
so true
@@LemonCake101 gonna use the link cuz you are literally me (schizophrenic autist)
ryan_gosling.jpg
Maybe EU IV AI is bad, but at least it is not CIV VI AI. In that game you set the AI to the highest difficulty and when you get to the modern era or so you look at the strenght of that huge empire and they have a whoping 3 obsolete military units or so. You can literally conquer them with no problems with them being 1-2 technology eras ahead of you. The only time AI is difficult is at the start since they start with like 3x your units plus get extra cheats(to compensate for them being terrible I guess).
While I don't disagree that things can get worse, I would still like to hope they get better.
@@LemonCake101 Ai peaked in the early 2000s. it takes you hundreds of hours of experience to beat the WC3 or UT2004 Bots on the highest settings
@@nox5555 AoE2 AI was a nightmare on high settings.
@@mircoss7217 I still remember the company of heros AI on expert.
The carnage was extensive.
@@JakeBaldwin1I still play Spearhead for COH2 against Elite AI with friends from time to time. The amount of swearing and anger from 4 young men over the internet is honestly impressive.
I did not anticipate lemon cake being a machine learning researcher but that 100% fits
shh I started it before chat gpt and stuff made it cool ok
also yeah i'm one of the people who play eu4 on very easy 😭i've played a few playthroughs with minor nations on normal/hard but atp i don't want to micro my wars too hard
@@loganstewart1161 hey man people play video games to have fun. If you are having fun, you are doing it right, and this is a hill I will die on
@@LemonCake101 unfathomably based opinion
@@LemonCake101 God Like opinion
It would be trivial to make an AI that can beat any player 100% of the time. Just make every non-player country no-cb the player on the 11th of december 1444.
Shh
Nah, I'd win
@@doudline2662 (type the console command "god")
Who's laughing now ?
son of a bitch, he really did!
@@Duke_of_LorraineYou can still get stacked wipe with that command
here's the thing: most people don't actually want perfect god-like AI, they want a challenge that they can overcome. The goals of default strategy AI are to 1: to perform well, even with a game populated entirely by AI. 2: represent a challenge to any players. 3: be a competent ally, or have coherent diplomatic strategy so as to not frustrate the player (unless backstabbing is the point). 4: not be unrelentingly ruthlessly impossible to beat, leaving no opportunities for a player to actually play the game. And there's even a 5th one in historical games like eu4: relatively likely to actually do historical things, or do unicorn/easter eggs occasionally.
This really isn't as simple as "make all 300 nations have perfect behavior."
No indeed, I do… talk about exactly that in the conclusion.
Sure, I don’t want perfect. I can’t best perfect lol. However, I wouldn’t mind an AI that isn’t get the human centered, and makes some sort of sense.
If you have a god AI it means you can tune it down to fit any difficulty. But if you have a bad ai, it will never compete with a decent player
To be fair, I would really love a "historical focus off" button in eu4. I agree that the AI should be more historical in general. I had several problems seeing Portugal AI colonizing Brazil, which is literally their main focus...Also, I have never seen GB go for India.
But, as I already said, if we go more historical, I would also like an option that goes less historical. Just like Hoi4 has this button.
Chapters in a sponsored video is unfathomably based
Glad you appreciate it!
I thought you mean “Why AI sucks” and was going to be lectured on this rhetoric for 27 minutes
Dw you can put this one on the CV it has sources and hence counts as research
As a Hoi4 player i find the AI frustrating in different ways but mainly working with the plane/ship and tank designers as it will make the worst designs imaginable which means that with the right design you can beat it 1:20 in terms of production cost (when using planes as an example). This is dumb and i think i would feel a lot more challenged and have more fun if you could exploit the ai so easily
I remember loosing a ironman run in Stellaris when a AI gave me a ring world that drained all my budget.
Clever tactic from AI!
”It is a trap!”
😂
Depending on how much of processing time ai decisiinmaking of all hundreds of nations takes increasing its complexity might also seriously slow down the game
Indeed, that is something thats always worth keeping in mind!
I think to improve performance there should be two gameplay levels, GS (macro) and RTS (micro).
The RTS level is more efficient but takes more time and computing resources. Players and AIs directly involved (neighbours, vassals, enemies in war) with them should play on this level.
The distant AIs should instead play on GS level. If I'm playing as Albania or Switzerland (or even Great Britain), a battle (or even a war) happening between Ming and Oirat should be something neither myself nor my computer should be concerned with. In such a case, instead of using RNG for every single battle there could be a single randomised calculation for the whole war. Instead of having the AI develop each province separately or build every single building, there could be a whole area development. We could even completely remove computations for distant AIs by having predefined behaviour for them. Distant in this case though would mean that they and the player are not affected by each other's actions in no way whatsoever. The switch from no computation to GS level computation could happen when you get close physically or when you interact with any of their neighbours.
In the first versions of the game there were much less provinces, but this way there could be way more. It's just that the majority of them would be dormant for a huge part of the game.
Something I think the EU series could take advantage from the CK series is using internal vassals. Yes, you have a much smaller demesne and less direct armies, but you would also use less time micromanaging and your computer could perform better, especially when you have a very big country.
Yeah but that number is gonna reduce dramatically as countries annex countries
Which is why mods like Starnet are absolutely essential when playing Stellaris. Unless you want to fight authoritarians because Starnet gives full citizenship to everyone regardless of ethics so it just feels like you’re fighting the evil oppressors but they treat their pops the exact same except their leaders wear shiny hats.
Of course it gives full citizenship - it's always the best option in the long run. There is no benefit to maintaining any kind of underclass, as all it does is make your downtrodden pops less productive. Which perhaps reflects reality, but it ignores the real-world political pressure of social elites, to maintain the status quo even if reform would be a net good for all.
Stellaris is full of these kinds of non-decisions, where there is no reason to take a short-term benefit at the cost of the long-term. Because Stellaris, like many other strategy games, is a progress race. The empire that grows fastest will inevitably dominate all of its peers. So the player is pushed to always prioritize long-term gains, to keep up their place in the progress race. Keeping the commoners in their place might make sense to a short-sighted elitist, but to a far-sighted player it's just a massive opportunity cost.
Crusader Kings avoids some of this, by keeping the player rooted in the present moment. The player has far less reason to focus on long-term goals, because the simulation is so chaotic that there's no guarantee that any long-term investments will pay off. So the player has to just focus on their current character, and play the game life-by-life, adapting as they go. It's not perfect - the player's ability to Quantum Leap into their children produces dynasties with impossibly consistent long-term goals. But it's still vastly better at producing a historical narrative born of short-term decisions, than other strategy games.
Except that Starnet hasn't really been update to support new game mechanics/changes for years since the original mod author is from Ukraine and can't update the mod (it's only update to run on newer version but the AI hasn't had any change for a long time).
So my paranoia of the multiple moments i sent my armies to stop a siege from inside, what would be, the AI fog of war for then the AI stop the siege as soon as i unpause and proceed to manouver all around the entire mediterranean sea was not the AI cheating but the AI being successfully dumb enough to bait my army and then plant the seeds of confusion on me about their army location. 10/10 AI
The "AI cheats with fort zone of control" -cope probably stems from the fact that the ZoC rules themselves are at least a bit convoluted, but provinces have very irregular shapes and formations that can further lead to confusion. Worst offender is one province in Burma, that is so small and sits so snugly next to the Ava's capital level 3 fort, it literally is inaccessible until the said fort is sieged down. This effectively grants the AI a safespot where to rebuild armies in. I'd argue that the patch where AI went way overboard in deleting forts made the 'The First Toungoo Empire' -achievement easier by an order of the magnitude.
I just want 5-10 AI personalities instead of three levels of aggression. There is absolutely nothing wrong with a wildcard AI either. Civ Gandhi's kindness overflowing into 11 out of 10 aggression is a great example of a community appreciating such.
That's an urban legend, actually, the programmers have denied it in official interviews
7:38 the ai in ck3 is fine until a crusade happens in which it will become insanely brain dead and will play like rodents running away from winning battles and baiting you like rodents.
Crusades are mind numbingly stupid in ck3. We had all of Israel captured, and what does the AI do? Surrender! Like wtf we had them on the ropes. Tons of gold on the treasury and we walk away with nothing.
Imagine every country around you played by an AI that was trained on Florry's vods.
0:21 You're not suppose to do that? Big number go brrrr is how I do warfare
i cannot fathom their being better strategies
Looking at how they develop their provinces, yes.
The way I develop them, is specialise them either into money-provinces (good trade goods : workshop manufactory trade and mainly dev production) and manpower-provinces (weaker trade goods : barrack training fields... and mainly dev manpower). Meanwhile the AI doesn't seem to focus their provinces. This means that at equal dev and money spend, my provinces are overall more effective as everything synergises nicely, while for the AI it all looks random and unfocused.
So ironically they can focus them! They are literally coded to not do that though, and maintain an 'even' distribution
Your way is bad as well though
your strategy only "works" precisely because the AI is bad. If it were competent, the population of the military provinces would decrease over the years, as everyone would migrate to the developed provinces, since no one wants to be condemned to spend the rest of their life in a miserable place until they die in a war. In a few years, you simply wouldn't have anyone left to recruit into your army.
@@chombus2602 that's the fault of the game systems, having this abstract development instead of a proper population mechanism.
That is why I always play on at least Hard.
AI doesn't know how to specialize its resources well so at least I can balance it by giving it flat buffs.
Somewhat related: One of the many things holding back the 1999 'Sid Meier’s Alpha Centauri' is that the AI isn't that good. (More imminent one is the archaic UI and control scheme.)
However, someone raised a point about that this isn't because 1999 technology didn't allow good AI - the AI in more recent Civilization games isn't any more advanced. They've just started designing the games around the fact that AI is bad at tasks like long-term planning and managing very limited resources. Aside the main series never having some of the SMAC features (unit designer, orbital infrastructure, dynamic terrain/weather) the following were removed after Civ III (that was developed somewhat concurrently with SMAC which used heavily modified Civ II engine.)
Civilization IV removed the city-based supply (have to pay shields/minerals for unit upkeep every turn) in favor of global unit pool with certain number of "free units" and rest are paid with gold. Wealthy-Science-Luxury-allocation was also removed in favour of a simple slider on how much (0-100%) of your gold income you invest into science.
Civilization V completely divorced Science and gold production. Units now always have their per turn gold maintenance cost.
To bring this comment back to the Paradox games, it could be that Victoria 3 trying to remove the traditional army stacks was at least partially motivated by this principle. AI playing the game "wrong" makes for an extremely frustrating experience (either your ally is useless or the enemy army marches across the entire Eurasia to either run away from you or siege a remote region.)
I wouldn’t really say Vic3 was made with improving AI in mind, it still screws up in wars and mismanages economy to the point they get easily eclipsed by the player sometimes more than a hundred fold
@@newguy8288Depends on the situation though. I actually have a harder time against the AI in Vic 3 than most other Paradox games simple because they love to join the opposite side of a war.
SMAC was such a fun innovative game. I feel like dragging that out again.
@@newguy8288 I tried to argue that instead of making the AI better, the game is made dumber so that the AI can handle it.
@@Xazamas in the case for Civ 5 I can say that is correct, but smth like Vic3 is a mixed bag for me
I recall playing a strategy game from way back called "Heirs to the Throne" and it was notorious for having an unforgivingly strong AI. There were no settings to adjust the AI strength, but the rules of the game could be changed at creation. The AI however never made a mistake and very carefully read its odds and even under favourable starting conditions, was immensely challenging to fight, and this is with only one active. The game supported up to three AI opponents as well as the unaligned forces.
After spending way too much time working on AI for a city builder, I don't judge too much when game AI sucks.
It‘s not even just battles and war strategy, it’s how they are so unable to administrate their country without buffs from hard mode
In case of Stellaris and HoI4 the death of AI was easy to observe as both games had evolved too far from original framework. Especially in case of Stellaris. It's like trying to make Counter Strike AI play Dota 2 just because they're on Source. And I can't suggest a counter to this as evolution of gameplay was in general a good thing.
The issue was the shift from a tile system to the convoluted district/building system we have now, which has two parallel systems doing essentially the same thing at the same time.
As a total war player I just want CA to make a system were the AI has a rock,paper scissor system that says to the AI that charging all your cavalry into a pike wall might not be a good idea or maybe force the AI to actually use formations and make the formations used unique to certain factions or change them based on what the AI army has and what the player has and the formation the player is using because why is 40 units just charging my spear unit and allowing my archer tower to get 300+ kills
Add just STANDING STILL while being erased from existence by artillery. Problem is not that they let us win - problem is how stupidly easy it is to abuse them instead of outplaying them.
AI need to learn counters in like every game because for some reason a lot of AI is made more difficult to beat by just giving them insanely dumb buffs even in all the total war games, and it really should just become smarter, AI would be so much harder if it did know to counter cavalry with pikes and pikes with archers and everything else instead of it just getting +1 billion money and 15 armies of elite troops with +50 attack and defense.
I will say, the god awful AI of HOI 4 has been mostly fixed with the Gotterdammerung dlc. While certain aspects are completely broken (fleet submarines evaporate everything. There's basically no reason to build any other ship once you have the tech), the army ai has improved significantly and actually moves armies by concentrating forces at key points to attack. Makes for more dynamic gameplay and actually challenging at points, even as a major power. I wish the different difficulty modes affected ai behavior though, instead of just giving ai buffs
Nah, the hoi4 ai is still brainless, just a tiny bit less now
It still cannot design even half decent designs for tanks,planes,navy or templates, it also cannot really use MIO's properly
@@RobbiusBossius True. At least it doesnt give up instantly once you build level 10 forts and just sits there.
Well from the game developer pov, "smartness" of the AI is only one of the concerns.
The most important one, being performance of hundreds of individual AI actors making their calculations in real time. Because some advance stuff like learning models, require a lot of power to run even one of them.
The second concern is, that's you can't make AI, to independent. It's seems counterintuitive, but complicated and independent AI algorithm without a clear boundaries is a pain to debug and fix or even change. What if AI will find that's the best war winning strategy is split and merge it's troops at the speed of frame rate or even delta timez which gonna make battle eternal? Or even just crash the game?
Well and the third concerns, is that's from the gamedev pov, AI goal is not to win, or be smart, but to entertain the player and provide illusion of smartness or challenge. Some objectively better strategies, just isn't fun for the player.
One thing that always bothers me in eu4 is the ai being almost completely immune to land attrition.
I’ve gone out of my way to avoid going over supply limits, stacking attrition loss reductions, etc. And I still take 10x-100x the amount of losses from attrition that the ai does. Meanwhile they’re siege a fort in Siberia with a 100+ size army or marching doom stacks everywhere.
In my Japan game, in a massive war vs the Ottomans, they marched 2 60 stacks into my Siberian territory. I took more attrition than them
Thats because AI can't handle attrition. Paradox very early on decided that AI is able to cheat at some things, otherwise it just breaks. Military Access for example. The AI doesn't use diplo slots for military access. Why? Because it can't handle it properly and will destroy itself with diplo points or not be able to wage war 90% of the time. (its also good for the player, since you can move with their military access and dont need to use diplo slots for it)
@OlafJorigson It felt appropriate to mention since it was brought up at 20:48 they don’t take naval attrition but didn’t mention land attrition. It may be unlikely to lose a war because the ai ships didn’t take an extra 5% damage at sea, but I can easily see a war being lost because I’m taking an extra 50%-100% manpower due to attrition just from sieging castles.
I once had a war between me + 2 ai vs 2 ai. I let the 4 ai fight each other while I just captured castles and avoided any fights. I still took just as much casualties from siege attrition that each ai lost (individually) to battles.
I would argue that the arguement at 5:30 is wrong. I haven't played rocket league but with convolution and an autoencoders it shouldn't be hard to compress a complex 3d environment down to a vector which can then be used as input to an NN. Action games environments aren't magnitudes more complex than grand strategy game environments.
Yes but they are infinitely more random and difficult to predict. Whereas in PDX games everything can be solved with a simple weight on a decision
An easier example is hoi4: there are mods that just add a bunch of triggers for the AI thanks to which it becomes almost as good as an experienced player, and even better than Xorme when it comes to differences between vanilla and mod. You get to fight blursed builds like 100 div mass mob Poland, ~15 heavy TD divs france, or a meatwall of 10mil red army troops in 41. All that with a few megabytes of code (probably even less)
Now try getting anywhere with a wall of "if"s in Rocket League without it taking up 200 yottabites of storage. It's much more difficult to make a good action game AI, and having to use a NN as you said is probably proof of that
@@maciekGTR I don't think you know much about machine learning... You can look up function approximation methods for reinforcement learning.
From what I have observed, rocket league environment can really be described as a finite list of elements in a finitely sized 4D space, where each element only has finite parameters. You can easily compress this down to maybe a few billions states. It's gonna use a few PB space at most. And then rocket league has a smaller action space then EU4 too.
Shame not to cover the AI in HOI4. Definitely a contentious one. The recent force concentration changes are amazing but it still leaves a lot to be desired, and frankly in HOI4 I'm 90% sure the AI can cheat.
I did have Hoi4 in the script but I literally cut it out since well even in current form the video had 6.5k words… viccy didn’t make it either
Because it isnt stacking +65% Transport shit combat ability
Transports are too OP already
I choose to believe this is a lapsus.
I wish you could stack TCA that high. But you can't even really stack it, because there is only one modifier of 40% TCA in national ideas. 😢
And only for RNW nations or custom nations.
love the measured, fast narration and focused, sober visual aids. too many people do "streamer voice" or make "funny exaggerations" every other sentence, fill the screen with wacky memes etc
Major powers breathing down your neck isnt just a roleplay element of playing as the Dutch. It is the defining characteristic of the Dutch in this time period. You thought we went all the way around the world for fun?
Anyway, I think there is a lot of improvement to be made in games using neural networks for the AI, just to make it play more like a human player. Train the easy AI by having it play against Bob the intern with 100 hours in map games and train the expert AI against people who have nolifed map games for 10K+ hours.
No indeed, but I said it as moreso that in especially older Eu4 versions, the Dutch land was so good and the AI was so incompetent, well, you became the Major really quickly
@@LemonCake101 That actually happened in real history, though
If you don't want anyone actually playing the game that's what you'd do, because a neural network AI would exploit the shit out of every mechanic in ways that players can't.
@@Gustav_Kuriga Well, considering its a game you can pause at any time and is semi-real time, semi-turn based, there isnt anything the AI can exploit that people cant. You could theoretically pause the game every tick and give micro commands to every individual army unit.
Of course they AI will have to get some limitations on how much effort it puts into the game. But the AI that is only trained to beat Bob the intern wont need to exploit much to do so. The expert AI might learn some exploits, if that becomes an issue you can just forbid it from doing those actions.
But players use exploits too, so i dont really see your problem here tbh
@TheSuperappelflap You do realize games can have bugs/exploits that only appear when you input commands at thousands of times per second right? Do I really need to explain that to you?
20:53 oh that's why i still don't have the first circumnavigation achievement
Absolute Cinema. Didnt think cakeman woukd drop a PhD thesis on eu4 bots.
An interesting thing with Paradox sieges in particular is how often the mechanic gets re-used. The mechanic is "Roll 1d14. Add modifiers. If the value is greater than X, complete the siege. If the value is greater than Y, add +2 to the stacking modifier value. If the value is greater than Z, add +1 to the stacking modifier value." But that EXACT system gets re-used all over the place. It's the system used for First Contact and Archeology in Stellaris. And they ripped it out of the Europa Universalis board game. EU4 is just the first time they revealed what was actually going on under the hood. At least that was according to one of the people in charge early on in EU4's life cycle (pretty sure that was before the Conclave DLC for CK2, but I might be wrong). Given that mechanic, it probably shouldn't be too hard to train an AI how to evaluate sieges. Especially if you're factoring in "Is it worth losing the modifier I have in order to go do this other thing, and potentially come back with a bigger modifier in this other area?" Though you would also need to factor in time and such. And you could probably use that for the AI to evaluate whether or not a siege is actually worth attempting with the forces it has available. I don't know the current modifiers used in EU4, but I suspect that training problem is much simpler to approach than the full game. And having a neural network work out a system that would work for those evaluations, with minimal changes to the codebase, would probably lead to some helpful patches. I don't know about the rest, though.
I think another issue with the Stellairs A.I. is that over the years more and more stuff was added to the game that rewards planning ahead and understanding how things work together. Things that a player can effectively use very well but a A.I. struggles to even begin to under stand. The biggest sign of this defeat is the "difficulty scaling" option, which really is just a band-aid on a heavily bleeding wound.
Machine learning algorithms are incredibly computationally expensive to ship into everyone’s games, yes sure you can train it beforehand and ship a set of weights to the players instance but Therein lies the major issue. The ai requires hundreds of unique “thinking equations” which determine the mathematical value of each potential action, then it must choose the highest valued choice. Eu4 has hundreds if not thousands of game state variables so you can imagine it’s hard to formulate these equations and add weighting on top of them in a balanced manner. The equations also have to be consistent between themselves ie if the think equation for building a granary naturally increases faster than the one for building a new unit it will be skewed. This is from my own personal game dev experience, I cannot imagine how hard it must be to make these for eu4. A challenge to you lemon, try to formulate an equation for each potential ai action to gain a perspective on the developmental difficulty of this
One of the problems is the lack of communication between the player and the bot.
I've had this issue as a hoi4 player, where the AI covers fronts I've already taken care of but doesn't care about the ones I cannot spare soldiers to defend.
Im glad someone pointed out that it's not really AI and it's all just machine learning. I can't tell you how annoying it is when you hear all the media, Nvidia, CEOs, and youtubers going around talking about "AI".
I never really saw that AI fort problem in EU4, the path finding and making decent fort lines that actually lock down an area is pretty easy to understand for me.
it's basically "it's bad to make player feel bad at game" combined with some performance considerations. Developers could build AI that is at competetive level or near it - especially with good micro. Just look at AoE2 HD edition hardest AI. It's quite strong on competetive level and this game is also very complex technically
I had CK3 AI within life of a single character conquer almost the entirety of europe except Iberia and british isles. Some random swedish guy got Great Conqueror and Scourge of the Gods.
I always argue that the issue with games, EU4 especially, isn't that the ai is bad but that the games have been built up to be too easy or too boring internally. It frustrates me so much that EU4 ai is sometimes megablobing, like getting the russian 1910 borders in 1600, because I then have to blob hard to keep up. I want nothing more than getting the french borders in 1789 to be something that is both hard and satisfactory to have done. Its so annoying that Worlds conquest by 1600 is commonplace.
Yes, and I hope EU5 will be more like you said. Well they said that you should blob less but we'll see
The current AI is good enought for people who play on very easy or easy. But for people who want to face someone with some competence at the game, it just fail to do so, it is almost as it is programmed by someone who isn't particularly good at the game and just wants the AI to do stuff that doesn't look like a bad decision at first glance without the context of the situation.
In my opinion, if you want to go against competence in a strategy game, you shouldn't be complaining about AI, you should be playing other players.
@@Gustav_Kuriga If only it was easy to setup.
23:37 you forgot the planets dedicated to purge xenos!
I had a run on shogun 2 as the Takeda. Was very early in the game (less than 20 turns in). An enemy clan marches a full stack and guns it for my territory. I was forced to basically reload just to prepare to fight this one army by scraping every man I could and forcing a battle which I could gain every advantage I could. I was against some really good troops and lost once or twice. I had to reload as if I lost, my campaign was over. I was able to finally win a pyrrhic victory and the ai army retreated. Although I had to save scum, it was EXHILARATING to experience and overcome the challenge.
Now, would I want that every 10 turns? No! The ai has to be Stupid sometimes so the player doesn't get absolutely stomped. It should also be smart sometimes. I think its hard to make a scenario that lets the ai be smart without removing lower skilled players from being able to enjoy the game
Shogun 2 AI was good in that it simply was aggresive. Did well to consolidate late game (lowering endgame turn times), before Fall of Samurai had workable naval AI (that simply attacked thus putting player instantly in pressured situation) and was kindly offering player epic castle defences (meat grinders) that were still something you had to work through but were satysfying :).
yeah my immediate thought I had here was with modern machine learning they absolutely can make the ai the best player to ever play the game
24:15 As a stellaris player, the AI building random stuff on its planets is super annoying and slows down the game a lot
That deck of cards thing would ultimately push some crazy players into waiting the bad thing they don't want to be played before they engage in important battles :')
To add Vic 2 to the mix, they AI in that game is painfully awful at waging war. They will randomly seige provinces early game when the first priority should be army concentration destroying enemy armies, and late game it is so easy to lure it into battles. Like, you put a small army on a mountain tile, the AI will attack it, then you reinforce it and win, and when you attack, the AI won't have a backline to reinforce. The AI also leaves provinces open late game even if they have a large enough army to create a frontline. Outside of fighting wars, the country AI seems fine, but man, do capitalists suck at being capitalists. They lack the ability to optimise production like in Stellaris
6:23 It’s the first time I’ve heard the Māori mentioned and I’m offended it’s like this
Congrats on 30k dude
In Stellaris the AI sucks because it doesn't know I am minmaxxing with 6 word documents everysingle day that passes.
Also they don't plan ahead, only if there's a sudden unemployed pop or a deficit of some sort it starts to do something about it. Otherwise is like: "I sleep."
For the dice roll randomness, the Fire Emblem series has had a number of neat math tricks to make it closer to how humans think. The most basic from the GBA games was just you random roll twice then average the results for the true number. With how it uses the rolls, this makes high percentages (90+) nearly always hit as the average is below it, and low percentages (5%) happen extremely rarely. Some of the more recent games have this formula become a bit more complicated, but it still shows how you can make randomness work closer to how humans expect it without doing the "Karmic Dice" approach which just end roll streaks.
That's very fair, and indeed humans forget that a 5% chance happens once every 20 times: so 1 finger represents 5% of their fingers and toes added up (helps with visualization I find)
@@blazewardog you're as old as comrade2020
@@blazewardog so Your subskrabt not even close
Playing Stellaris Evolved (a mod) made me realize how much they improved the AI. Actually letting them specilize planets (sometimes) lmao.
2:08 most based descriptivist linguist I've ever met.
Definition by democracy
While I agree with a lot of what you said, I think your argument against why AI can't be taught how to be good at grand strategy games is missing a major element - the scheming or plotting or even just planning. You could also call this 'the human element' or just strategy, but AI struggles with both judging their opponent and thinking ahead. You touched on it briefly with the player knowing that another player would have 300k more soldiers hidden, but didn't really elaborate.
Ultimately, if they made an AI that was exactly at my skill level, I would still win. Not because I'd just be better than the AI, but because I can lie, trick, bluff, deceive etc. an AI in a way I cannot do to a player and an AI cannot do to me. Very few AIs for video games have long reaching planning capabilities, instead being mostly reactive to the current scenario, and fewer still have a memory or the appropriate reactions to use it.
For a stupidly basic example, in HOI4 if you ever find yourself up against the AI in an unbeatable position, often your best choice is retreat. Not because you've been beaten - but because if I fall back, the AI will see "oh, open tiles!" and walk into them, claiming them. It will move its entire force with the frontline and leave no-one guarding the unbeatable position. Then, I counter-attack quickly, and while his divisions are low on org and retreating I attack the unbeatable position again, now guarded by the broken remnants of his retreating force. A player would (almost) never fall for this, but the AI would. And sure, you could have the AI not advance into suddenly empty tiles - but what if I ACTUALLY retreated because there was an invasion on the other side of my country, and advancing was the correct move? The AI still has no idea where my units are, but a player could put 2 and 2 together in a way the AI can't. A player knows someone who was just winning suddenly abandoning their frontline and disappearing is a suspicious act, an AI just sees it as no longer knowing where those units are.
It's incredibly difficult to code an AI that can handle something that a player would do naturally, and I would say that there's no AI in these sorts of games I can think of that is immune to these tricks. That's why I find Chess a poor argument for why AI for PDX games is hard. Chess is easy for a computer to win, it's very rigid in terms of what each player can do on their turn, each turn is in order and each turn changes the possible outcomes of the next turn.
I code AI for my mods for ArmA myself and yeah, it really is just weights and decision trees all the way down, but a lot of my design work is around making the AI behave the way a player would, rather than trying to make it outright "good", and so much of it has to be faked with subtle cheats or forced behavior of the AI in order to achieve what looks like 'normal gameplay' to the players.
That’s also a very fair point. In many cases indeed I was focusing on mechanics but yes teaching an AI ‘strategy’ isn’t a trivial task either.
For stellaris in particular, I think it's that the AI gets more short term benefit from stuff like strategic resources so it goes hard in on that and ignores everything else.
3:54
If something can show that they can make a good AI (as CA showed us a decade ago) is funnily enough a different thing with the initials AI
Making a good AI for a horror game is very much a harder challenge that doing so for a “spreed sheet simulator”
Indeed!
I've long been curious how machine learning AI would play games like this or Dominions. Would be a lot to learn from such a thing. Good luck getting the resources to train one sufficiently and update that with patches, though.
Because AI does not play Lithuania. Day 2 of asking for a 1h+ Lithuania tutorial
Step one: Get integrated into the Polish-Lithuanian Commonwealth
My biggest Problem with most grand strategy game AI's is that they don't scale with the player. Individualy, but also over the course of the game.
Which leads to hard early games and easy late games in nearly all cases.
Total War is a great example of this. Early game on higher difficulties is really hard, because AIs on the other side of the map will drop whatever they are doing and send dedicated death squads to the player wihtout ryhme or reason. While late game is really easy, because the AI does not dynamically build up anything to stop your snowball.
In short, the AI does not adapt its actions to the strength or playstyle of the player.
Which feels like it should be really easy to fix:
* roll one or multiple random rivals and cheat for them to grow at a similar rate as the player.
* adapt these rivals to your economic and/or military rating
18:46 I do, just cause I hate the random 9-9-9 in important battles. Like, that is just completely random and if it's a big battle, tough luck, nothing you can do, which can be frustrating instead of being able to see: "Hm I went in and they had reinforcements waiting, that I couldn't see. I should play more defensive."
I recall there being an issue when zone of control was introduced (eu4) where if the AI could find a land route to a province, it could move to that province and ignore zone of control. I definitely witnessed this issue watching the ai walk right through barrois to approach paris (metz -> barrois -> champagne) instead of going the long way through belgium that any player would have had to do. I would have confidently said that the ai still does this behavior, but thinking about it I haven't seen them do it and i've been playing a lot of 1.35.6 recently. I wonder when they changed that.
Meanwhile the ai absolutely knows exactly where my drilling armies are at all times because the moment I click drill because I don't care about a war i've been dragged into a 30 stack of random germans manifests on a fort-less border with their infinite military access to beeline at my 0 morale army.
22:46
that is probably one of the funniest things i have ever heard. Ty for this
On the deep sea atrittion thing I could see a situation where you have your army on boats run into an enemy fleet slightly weaker than your fleet but you still don't wanna loose ships fighting so you kite the ai fleet ending up in deep water and after a few turns they catch you and you lose the fleet and a nessicary chunk of your army snowballing into losing the campaign
The fundamental problem is that the initial statement of devs being just not good enough is true most game developers dont understand their game even to the point where something that is clearly op or up does not get balanced for literal years and dozens of updates. Honestly the easiest thing they could just do is make the interface that allows people to code and or train their own ais.
I thought Stellaris’s “AI” worked pretty well with the old tile system and admin caps. It would still do wonky things but you couldn’t outcompete it anywhere near what you can now. The amount of min maxing you can do is just insane.
An interesting factor in “AI” is how many opportunities and options does a human have to outcompete it?
The thing I hate about AI’s in games, is as soon as you ally them they become brain dead. Its like they literally break, or they’ll never ally with you despite being surrounded by 1 million troops of mine and only having their king left alive. Also hate when they run their troops around and never engage in battle, they just back cap you and flee etc.
They don’t have to be as good as players, they just have to “roleplay” the nation they’re taking over. With some variability so its not always the same build every time.
Stellaris ai planets are so bad, that I not unfrequently research planet cracking just to blow them up rather then fix them
It’s funny but I find the AI is EU4 is pretty competent it’s just that the players threshold for doing everything well is just much higher.
On VH a big austria is a massive pain. Anytime you leave anything short of an 80k stack somewhere they will come and hunt down your armies.
A couple things with the AI in Stellaris.
First of all the sector management has been fixed, one of the basic settings (that should always be turned on) with the management is to build according to planetary designation. If you set the planet designation to mining world the AI will build mining stuff.
Second of all in slight defence of the AI not specializing it's worlds. A player is very capable of adapting to circumstances and fast, but say an AI loses a primary mining world it's economy would be utterly crippled and it'd have deep difficulties fixing that if you tell it to specialize it's worlds.
Oh I wasn’t aware of those basic settings no I had them all off XD maybe those settings should be ‘opt out’.
Second of all in slight defence of the AI not specializing it's worlds. A player is very capable of adapting to circumstances and fast, but say an AI loses a primary mining world it's economy would be utterly crippled and it'd have deep difficulties fixing that if you tell it to specialize it's worlds.
Especially since the AI is incapable of demolishing anything.
Every AI Planet I take over, has like 100% crime
The AI can be brainless, however in an older update of Hoi4 there must’ve been hidden stats because I played exactly like the AI in historical for 2 nations, yet lost both times against each other.
Literally from the 1st statement. How can you make it if you play they game.
1) In general, AI obeys zone of control rules but it also cheats them sometimes in a way that players can't.
2) devs can't create a CPU player that would beat 99% of players. For them to be able to do that they must be in 99 percentile of player base (which they aren't, they are casuals). Even with dedicated machine learning Alpha Go and Dota AI that were trained on millions of matches were then beaten by average players when strategies were established.
3) In CK AI indeed behaves more on its traits than surroundings. And it's really good.
My personal gripes with EU4 AI:
1) It tries to contain player from day 1.
2) It can see though FOW past 2 provinces.
3) Diplomacy.
19:18 that's called triangular distribution)) Also they've implemented that system of not repeating rolls in Baldur's gate 3.
21:18 with that I'm going to argue. Imagine how you get FOW in enemy territory. For 2 layers of provinces so does AI but after that it can see everything. Put a doom stack deep in your territory and move it around. AI would react to it and move its armies away. To make it obvious use speed buffs or cheats for instant movement.
1. Again... no it doesn't: I promise you. It doesn't.
2. The people that made deep blue where not 'smarter' then deep blue at chess: what kind of logic is this??
3. Here we agree
1. I don't really agree not going to lie, but its another (yet again) optics problem
2. It does not. Watch the section chief.
3. Yeah honestly no clue what it does there
@@LemonCake101 I am sure AI can see the armies because if i click on a province to engage enemy all the way from siberia, the enemy stack moves, despite being stationary before
@@LemonCake101 Not that I agree with the OP, but EU4 "AI" is just a bunch of if/else statements whereas Deep Blue is actual AI that recursively self-improves. They have virtually nothing in common.
@ that’s fair, but my rebuttal on point 2 is it’s still conceptually something that’s possible if that makes sense
@@LemonCake101 well, chess have an simplistic rule set. Where AI has extreme advantage over low "human RAM". Plus game developers (as you've pointed out) don't create an AI but a bot.
In games where rules are constantly tweaked from game to game human player has an advantage over AI.
With player containment I got an easy example: when you start as a small tag (can't rival majors) majors would ally your neighbouring OPMs, but as OPM you can't ally those majors 1st month.
On FOW, when you get time, try to set up an experiment (establish delineations before start).
Complexity of how many mechanics you stack into a game for sure can hurt a.i. in how it creates allot of moments for cheese to happen from interacting with anything they not patch enough or does fly under the radar. The problem with a.i in videogames is usual that it does not get updated or maintainted very well and players learn its weakness.
Yes, this has always bugged me! Seeing how well Alphastar does at Starcraft makes me wonder how incredible it would be if Firaxis unleashed something similar on the Civ franchise. It could be a big marketing gimmick too!
I do conceptually love the idea of ‘super strong AI’s’ but not due to modifier buffs, but AI ‘skill’
Will be interesting to see if zone of control gets more or less complicated in EUV
Thats a very good point frankly!
24:09 Actually planetary automation does not work anything like the AI, instead the only thing planetary automation does is build distics/buildings based on your designation whenever there are unemployed so on a mining world planetary automation only ever builds mining distirics or the building that impoves mining, that is as long as you turn off the option to checkbox that let's planetary automation build strategic resource buildings wherever it wants to build them. All you need to do is
1. Manually set your designation
2. Click the settings wheel icon on the planet automation thing and set the settings appropriately.
A potential technical issue isgame performance.
The ai is already one of the things slowling most pdx games down the most, so it being light on performance is one of the reasons it is so simplistic.
if Paradox can make the ai better then can they finally fix magicka 1 being super unstable?
Right after they release the Empire of Sins Expansion 3
Right? I am so sad about this, recently I upgraded to win10 from win7 and when I went to play it crashed like every 25-40 minutes? Such a good game, Such a shame really.
You don't mention most important thing: performance, it's possible to make good ai, but probably it will make games unplayable even on NASA pc. And CK ai win at this category too because it doesn't struggle with late game performance.
Performance does matter, but you can make a really bad AI with bad performance, and a good AI with good performance too. They may be connected, but not 'inherently'
I absolutely agree with first part, but don't think its possible to build a good ai with high performance, especially if you rely on ANN, it's just math.
@ if you do say so, I don’t know enough about that topic. But I have seen strategy game AI’s that do well in older games that run fine today
What I must say about AI and the rocket league thing is that AI is naturally gonna be balling when it has all the information like in chess and rocket league, when it doesn’t have all the information its a different ballpark.
Victoria 3 AI is pretty decent these days- diplomacy values are pretty straight forward and react to things that you would think they should react to and does its level best to explain them. AI is annoyingly good a military even if it doesn’t use meta templates and it can build in the economy well but not total meta
The only "AI Cheat" I feel is the lack of clarity in troop/ship limits or what not. For example, if you reduce Venice to only two provinces, they are still able to maintain a fairly large naval force limit. Secondly, I can feel confident infighting a nation that has 40k troops and 20k manpower, then declare war on them. However, theres always some variation in the amount of extra troops they can raise, and it can feel absurd when the AI has 40k troops, 20k in reserve but the next month their troop count jumps and with three mercs stacks to be at 72k. Like were they not near their force limit at all or is the AI that willing to go into debt?
There was a recent interview with a former Creative Assembly AI developer who claimed that he was told not to make AI too competent so players wouldn’t have a challenge, though take that with a grain of salt. I think best example of a strategy game with a competent AI is GalCiv 2.