It is partly because many FE devs like promi were already +10 years ago part of community making AI scripts, maps and campaigns (and etc.) before they started working on Forgotten Empires mod and become devs later.
"I want it to be as strong as possible on the hardest difficulty. I want it to be as fun as possible on every difficulty." Words game devs should live by.
@@AbdulGoodLooks he basically said that moderate AI is dumb because he feels they're not fun to play against because the way he likes to play doesn't work.
@@tejasaivuyyala TheViper has been trying to win against 4 DE extreme AI for months now. Any improvements made to the AI is unfortunate news to him and his plans to beat the extreme ai. XD
@subhadeep mandal The only map I've found for that to be the case is Black Forest, because it can be really hard to win against mass siege micro with no space to move units through. Other than that, getting a proper economy up and running and countering an attack or two will let you build up towards a ball of death in like an hour. As for hardest/extreme, I've found that the game is usually lost or won within the first 25 to 30 minutes, because you can either deal enough damage early to where it can't boom properly and beat it quickly or it wins one engagement into outproducing massively to where it's almost impossible to come back from.
Didnt ask him about the Viper vs extreme AI challenge? I'd be interested to hear him answer whether he uses those videos and strategies for ideas on how to improve the AI, since Viper is basically his tester trying to exploit his AI in various ways.
The first time I played against the extreme AI (AND WON, YEEHAW!) I was trying to show a friend kinda how to respond to the enemy aggression, quickwall, counter enemy units etc etc. I didn't expect, however, the AI to do a drush. Caught me offguard but managed to survive long enough to beat it. Man, it was so hard but felt so good!
I wonder if you could do something with having the bespoke AI leveraging a separate machine learning AI. So instead of having a machine learning AI that plays the game, you instead have a machine learning AI that answers difficult questions that the main AI might run into. "Is it time to attack now?" "Should I drop civilian pop to zero to try for an overwhelming push?" I guess that is probably best left for when the limits of the bespoke AI are being hit and its more or less locked in, and it sounds like he still has a fair amount of room in that regard.
The thing with ML models is that inferencing scales pretty well with model complexity but the initial hit of having that model is substantial. At the end of the day you're better off just doing the whole AI in ML.
As I understood it, the problem is getting the information out of and into the game. There is no interface for a ML AI to command the game. But even if you use ML only for certain decisions, i.e. only using the results of ML, you still need to get all the data out of the game for training. However... we do have recorded games. So you'd *only* need to convert a recorded game into something usable as input in a neural net. Which probably isn't simple, but actually quite more doable than an interface for the running game.
I think someone could "crack" a machine learning software into the game, but it only could learn about the decisions made at one game. I'm inclined to a different approach: a patch that allow players to be recorded, then send all the data for a match to Microsoft and they could use a system to analyze the data. They don't even need video recording, just record of clicks, coordinates and time. Also, the seed of the map and then feed it into an expert system returning a new AI script. This could be used as montly patches.
I think that would be really good. I'm not sure if other players would enjoy playing against an AI when they were trying to play against players, though. As long as they only play a few games per week, and the games are spread out, I suppose people might not complain.
@@Jay_Sullivan I think it would take a little bit for people to realize its an AI. They'd have to give it a unique name like any random player and throw a few out there. Obviously if people know its an AI there are exploits to beat it. it would only work if its a surprise. I'm just curious if like the "hardest" ai is at about the level of the average ranked player or something on the ladder.
@@McSniffle worked my way up from 900 to 1200 the recent months. On Arabia I think the extreme AI would be around 1050, I could actually kind of relate to what it was doing. On other maps not so much though
Nice interview. Just to add to the machine learning, despite it taking millions of rounds for the ai to understand how to win a game, one of the other advantages is that glitches and bugs are sometimes being exploited by the ai and thus being discovered.
I'd like to see more interviews of this format; maybe with the creator of the Barbarian AI, other AOE Devs, or even people in the modding community. Good stuff!
I agree that machine learning for the ai needs access to editing the main code of the game to program it to allow machine learning and develop its own commands and ai commands to evolve and get better. Although one of the things that the ai does very well since the original age of kings was for it to dodge the projectiles coming out of any large projectile making unit: mangonels, bombard cannons, and cannon galleons.
One feature I'd really like to see would be assigning different difficulty-levels to individual AIs, rather than all AIs being fixed on the same level.
THIS PLEASE! Its so necessary when playing against the AI with your friends who have very different levels of play... Most other games after aok got it.
I just hope they can finally fix the villagers bugging out. Every game I find the weird bug where 2 villagers bump into each other and keep doing a little dance for like 30 seconds. Or worse they just freeze.
If the AI would focus villagers in their rushes, it would be so much more difficult. Whenever they come in Feudal with 15 Archers they attack the first building they see. You can get away with just building no units in Feudal, going straight to Castle Age, and all they do is take down a house and a barracks, before you can engage with knights for example. This is the case also on extreme difficulty.
@@tomekt9675 I can easily beat all AI difficulties on HD but it's so chill vs easy AI so I always fall back to that ^^ Just add extra opponents if the difficulty is too boring. You end up fighting an army worthy of an extreme AI, but with the general ability of an easy going AI.
How does the AI avoid getting stuck with something? For example idling units for extended amounts of times, or running back and forth with villagers, constantly switching tasks. Is it all hard work to fix these things one by one, or is there a grand plan to solve these issues?
Wouldn’t it be cool to have an agent of each AI’s difficulty playing on the RM Ladder (even if it were only to play Arabia Random Civ). I’d be super curious to know where they stand
I think the note that the AI will know your army size and act depending on that, will be valuable info for TheSnek. Might influence his playstyle vs the 4 extreme AI.
My only real problem with the AOE AI is how bad it behaves in long games. I'm not an especially good player, but my cousin and I play a team game every once in a while where we take on multiple or even many AIs at once. However, since we both like the big battles for territory and are pretty vulnerable early game, we like to play with treaties. Or last time we played on the Michi Map so the AI couldn't meet up with us until we were ready for it. We usually battle the extreme AI because we like its behavior in battles and its fierceness. But even it has so many quirks that always take away something of the fun. And I don't think it's because the AI is bad or something, but just that 90-minute treaty games with 500 population are understandably not what the devs are focussing on, but it would be so nice to change up some minor things to greatly improve the experience for us casuals. First of all you really notice how it handles villager production and military in a standard 1v1 no-treaty way. It will always produce some amount of military units and stop booming with villagers at a certain point. Which is weird to me, since I've seen it delete units to train different ones before, so I don't think stockpiling villagers is such a big problem for its code. But having 90 minutes of basically free booming, and then wasting so many early ressources on completely useless military units, only to then stop making villagers after 100 or so when it could go up to 500 if it wanted is really weird and makes outbooming it so incredibly easy, that we most often decimate each AI until the last one, and then keep sending ressources to that last player to keep it fighting. That said, it's also not concerned with having too many useless villagers. Because even if there is absolutely no tree left on the map and nothing to farm, the AI *still* tries to maintain an "economy" of ~100 villagers, even if 80+ of them are just standing around aimlessly. They are not tasked, they don't build, they don't farm, they just rot away and steadily replace every villager that gets killed. What also happens is the opposite: The map still has plenty ressources, but the AI doesn't renew any camps or farm ressources with more than a couple of villagers. There could be huge piles of gold in its base, but if the game went on for 4 hours+ or so, the AI will just ignore it and have its villagers stand around aimlessly until they die and get replaced (with equally useless vils). This is really sad because I'm a huge fan of the game, but this kind of behavior makes the games still really boring because it first starts lagging due to all the useless military units running around with nothing to do (but still visiting each corner of the map and spreading out). And once you survive the initial flood of units and destroy part of their village, they either are out of ressources and just become useless, or they don't rebuild and die anyways. I feel like these issues are not even that hard to fix, they just need to be actually worked on because in the usual competitive 1v1 games, this will never be an issue, thus it will never get changed.
@@Whisper0ak what are you talking about? Did you read what I said? First of all, to actually answer your question: Yes, I wonder why you'd give up if you had a couple 100k ressources of any kind and 5 teammates to help you out, and you instantly become useless once some basic infrastructure is destroyed. I'm not talking about complete annihilation, I'm saying that the village is broken open and the TCs destroyed if you will. But if you float basically infinite ressources for more military or a complete rebuild of your base, and you have the space and allies to fall back on, what's the problem in doing so? Why even stockpile ressources in the first place then? But my declared issue was that they don't even need to be attacked or destroyed, but that they just stop acting altogether after enough time has passed. If *they* attack *you* and you destroy their army, and they still have ressources, why wouldn't they train more? If there's ressources on the map, why do you spam an insignificant number of villagers, only to have them be idle for the rest of the game? My whole point was that the AI becomes completely passive over time, no matter my interference. And considering it's the most difficult AI, it should always keep fighting if there's a reasonable chance to bring it back. Basically surrendering once your first wave of Offense died, with plenty ressources to spare, the extreme AI should not be so demoralized that it gives up. I seriously don't understand how I'm "being unreasonable" here...
@@Whisper0ak only that the AI hasn't "lost" if it still has the economy, ressources and military to win. That argumentation makes little sense. It's not really close to "modeling real players". Only because you have a very linear definition of how "real players" would play the game.
The way an ML algo could work is by properly generalizing events in game to higher levels and then generalizing the reactions so that the model has less avenues to consider. Chess works because you can break down every possible move into a (admittedly massive) tree of possibilities with each decision and it’s possible payoff calculated. Do the same for AoE and it’s possible.
It must be fun seeing this stuff beat Viper in a 1v4 and he's amazing at everything! Good work with Promi and I love your videos Spirt! I have watched every one!
When a game is played . It saves file . It can replay. We can select opposition team. What did AI did? How this file saves each players single activity. Please explain.
Spirit of the Law please make a video about building. Things like, how much hp do buildings have as soon as you place them? how many resources do you get back if you delete them before finishing (at different stages, if the building was damaged)?, how do using multiple villager speed up the build time? if I want to produce skirms urgently should I build two archery ranges with one vill each, or one first and then the other with both vills? etc.
One question that wasn't answered, is the AI capable of playing properly at different maximum population? For exemple, is the AI balanced for 500 population or is it made for 200 pop?
With or without machine learning, the hard AI beats me like its no man's business. I need to be scripted, not the AI :D Also, I would love to see Viper's reaction to the latest changes in the patch that include the improved pathing.
Would AOE be the most difficult game for an AI like alpha star to learn ? I can't think of an other game that has so many parameters, randomness, where macro is so much more important than micro, 30+ civs to learn, so many different types of maps to learn ...
@@lalramdinavarte4987 Starcraft has more emphasis on micro and less on macro than AoE. Starcraft has less map generations. Starcraft has less civs. Overall, there is probably less decisionmaking, which is always the really hard part for an AI. Pure micro with one strat can bring you really high on StarCraft, it will never work in AoE.
Exciting to see if there's going to be one too, the progress will be entertaining to watch. SOTL did mention long-sword man when It comes to the Alpha-aoe2, so I guess It's not gonna be good under the condition.
@@R3DT1D3 Definitely not. A MOBA is relatively straightforward and the fact there are 5 heroes to control does not change much for an AI. The real difficulty is to add a proper playstyle for each of the 100+ heroes, but laning is quite easy to understand for an AI and even complex lategame strat are a lot less complex than in an RTS. Simply the fact that you'll almost always walk in a straight line towards the enemy base or occasionally flank is a lot less complex than in AoE where you need to set a priority for each area/resource collection, to know where to castle, where to hit with archers, how to loop with your units to hit the enemy base without being seen, where to protect with spears when the enemy will scrush... And lategame when and where to raid with cav...
I think what would help to make it more human like is to possibly limit its APM and increase it per difficulty level, because I remember when TheViper was trying to defeat the AI, but had extreme levels of micro that were far superior than any pro level human. Also I think giving the AI some more human tactics would help as well. MAYBE even some meme strats lol
So my friend and I recently started playing the game, and I still have some questions about the options within the online multiplayer lobby. The options are AI, AI (CD), and AI (HD). The HD clearly is the most difficult to us, and a moderate HD tends to play about the same level as the Hard AI. Is the HD one receiving the live updates discussed in this video? It doesn’t sound like it from the way he talks about HD here. Can someone give some clarity as to which setting is getting the live updates, and what the other options were from? (AI/CD/HD)
I assume you're playing the definitive edition of AoE2. The "AI" (often called the DE AI) option is getting updates, not HD or CD. It's interesting that you say that the HD AI is harder than the DE AI. I've found the opposite to be true.
Hi! So, the three AI options are from the three versions of the game: The original (AI CD), the HD edition (AI HD) and the new DE edition (AI). ONLY the DE edition is getting live updates. The other options are there only for historical completeness (so people nostalgic for old versions can continue to play with the AI they are familiar with). The 'skill levels' of the AI editions are not directly comparable - for example, the HD version has some weird difficulty jumps which were smoothed out for the DE AI, and DE AI got an additional setting. However, if you put a DE and an HD edition AI on max difficulty and let them slog it out, you'll find the AI typically beats the AI (HD) pretty easily.
I know this is a niche thing to ask for, but could you do a video on water counters? Like what ships counter what, when it'd be better to use your land units to deal with water units and vice versa, etc
Lovely and intresting Interview. 😊 I always wonderd why it isn't possible to selectthe AI strength for ever, AI player in the game individuell. It would be really great for balancing if the human players aren't on the same level. 🤔
Regarding machine learning: I think it would be reasonable to implement ML for small tasks (individual combats maybe?) while leaving most of the AI to for/while/if/else statements.
The ai could be much stronger if they have that, sadly there is no one interested in a project of this type. And if is posible to add few ML behaviours and not broke the ai, its obvius that a full ml agent would be much stronger but would also take more time. Maybe Aoe2 Alphastar would take hundred of times more time than the Sc2 alphastar. At this point i dont know if it would be worth the efforts.
A couple of questions from my side, does the current DE AI (including Extreme AI) has the ability to strategically onager cut forest / trees to make way to enemy town? Secondly, does current DE AI know how to use attack ground on mangonals for micro? If answer is no, is it planned in future? Thanks
Given the Conquers AI ran on a 233mh Pentium, and the new AI takes far more computing speed and power, the original AI's ability given the limited resources is quite remarkable.
Technically knowing the location of an enemy is not cheating as long as the location was explored. Constantly querying the map for buildable areas, you can guess where the enemy is. It is like trying to build a castle on a hill but the silhouette comes up red so an enemy is already on the hill.
8:30 I think the step between moderate and hard is now big. I can beat moderate AI pretty comfortably on any map with any civilization but get my ass kicked by hard AI everyrime :-/
Machine Learning, a fascinating things to play with. I don't thinks its because the ''map changes everytime'', its because of number of 'model based reagents' are much higher than starcraft or DOTA2, a particular model handling takes 10 to 11 neural networks on DoTA and all AI has to do is control and command one character. In AoE its build, maintain, gather, and attack! 4 model based reagents are difficult to control on current PC configurations. So evn if they build it, it'll be only playable on supercomputer facilities.
yeah, I also think it's not, in reinforce learning the algorithm designed should be capable to detect environment dynamically already (which is already a difficult part), and an AOE game itself can divide into many parts more than you mentioned I think, and overhead of that NN cannot fit in PC and more importantly is extremely costly for both building that up and running it
I would say the AI knowing how many units you have so it knows when to attack is definitely a cheat. Also based on the video the AI is still suicidal, throwing units away to TC's, archers ect. That was HD AI's biggest weakness because you could just wall and castle up and they'd just suicide everything they had and eventually would just stop doing anything (I don't know if that last part is still the case).
I feel there is too big of a jump in difficulty between the Hard and Hardest AIs. Any ideas why? I can easily beat the Hard AI 100% of the time. But I've been yet to beat the Hardest even once.
I find that hard ai is quite slow to switch to counter your army, while hardest ai is very quick so you either anticipate his counter, or push hard to win the game before he can switch. Just my experience
The ai now is definitely better than it's ever been. That being said, I would love to see it act more like a real player at difficulties other than extreme. I like to play the ai on hardest when I'm practicing strategies for online because it's similar difficulty to paint against players my elo (~1000). It would be cool to see the ai on hard/hardest attempting advanced tactics/strategies without flawless execution. For example making scouts and making 1 or 2 too many farms then struggling to transition to archers later. Or occasionally trying a scout rush, drush, fc knights, etc. For me on open maps the only strategy the ai does is fast feudal with massive amounts of archers and skirms. At my level I can put produce them but they're also not smart so I just fc knights and win. They're really good at eco management, but not realistic about using civ bonuses, military strategies, etc. Would be cool to see the ai play like a player at my elo: attempting advanced strategies/good counters but imperfect execution.
It's weird because I usually play Standard or Moderate AI's and the HD seems to be much more aggressive and generally harder to handle than the DE one, although the HD seems to completely overlook building castles and mining stone in general.. Does that mean it gets harder as you go to hard/extreme level? I have a hard time finding the balance between getting anihilated and destroying ennemis super easily..
I guess what is needed to get a machine learning AI is some API to access a running game from the outside (similarly to what a human player can do), instead of a script running inside it. Then you can use whatever generic ML technology you want to play and improve your game.
In response to which AI difficulty step up seems to be the steepest - I feel like it's the Moderate to Hard jump. I've only played Single Player, because I can kind of keep up a build in the first 30 villages, but then things get really overwhelming to keep track of, really fast. I can crush the moderate AI, but then once the Hard AI gets going, I find I have no clue how the AI has a 50 unit army and comes and rushes my base, and I've got maybe 10 or 20 knights.
I love SotLs enthusiasm for Machine Learning AIs but it IS crazy complex and would take a lot more than just Promi working on it. Just making an AI that FUNCTIONS in any shape or form is horribly difficult.
this guy is doing god's work. this game will last forever because of his AI. I mean, HD's was already cool enough. This is just going above and beyond. i wish blizzard games got as much love for their AIs.
i just love developers like these that take time to talk openely with the community
It is partly because many FE devs like promi were already +10 years ago part of community making AI scripts, maps and campaigns (and etc.) before they started working on Forgotten Empires mod and become devs later.
@@Tempires which is amazing isn't it
truely a part of the community
Indeed!
*cries in Team Fortress 2 dev team*
And Spirit makes a good work in interviewing them. And also, this is good for the game. Makes the fanbase feel involved.
"I want it to be as strong as possible on the hardest difficulty. I want it to be as fun as possible on every difficulty." Words game devs should live by.
@transylvanian this is a strategy game. Sticking to a strategy that doesn't work and then complaining that its not fun is pretty dumb.
Then why play in moderate? Is someone bulling you for using the easier difficulties?
@@yevgenyiyeet8737 translvanian deleted his comment and i'm confused.
@@AbdulGoodLooks he basically said that moderate AI is dumb because he feels they're not fun to play against because the way he likes to play doesn't work.
@@yevgenyiyeet8737 thx
3:25 A poor farmer gets conscripted to the army, colorized
One of us, one of us, one of us!
you're in the army now, oooh oooh you're in the army, now
*Sees one dislike*
I see that TheViper's watched this already
Didn't get you...
@@tejasaivuyyala TheViper has been trying to win against 4 DE extreme AI for months now.
Any improvements made to the AI is unfortunate news to him and his plans to beat the extreme ai. XD
I think the AIs just snipe his streams and should be banned
Haha awesome comment😂😂🤣
@@nbrocx4216 can he cheat by playing aztecs or some other monk civ on arena or black forest?
*_"Are you winning son?"_*
Yes dad, I finally able to win against easiest AI in AoE2 HD and now I'm ready for definitive edition AI
Chubbyemu meta lol
Me in a nutshell XD
I played onlinde once, I quit when i saw thre men at arms in my base befre I clicked Feudal Age :p
@subhadeep mandal Yeah, but it was difficult because it used cheats ^^
Now it's difficult because the AI is good XD
@subhadeep mandal I don't have DE, and I don't play hard difficulty in HD so I can't tell ^^'
@subhadeep mandal The only map I've found for that to be the case is Black Forest, because it can be really hard to win against mass siege micro with no space to move units through. Other than that, getting a proper economy up and running and countering an attack or two will let you build up towards a ball of death in like an hour.
As for hardest/extreme, I've found that the game is usually lost or won within the first 25 to 30 minutes, because you can either deal enough damage early to where it can't boom properly and beat it quickly or it wins one engagement into outproducing massively to where it's almost impossible to come back from.
a new SotL video uploaded 1 minute ago you say?
DROP EVERYTHING! I MUST WATCH THIS!
Didnt ask him about the Viper vs extreme AI challenge? I'd be interested to hear him answer whether he uses those videos and strategies for ideas on how to improve the AI, since Viper is basically his tester trying to exploit his AI in various ways.
At 3:26 the army drags the villager away
Good ol' peasant levies
Conscription explained in 3 seconds.
“No I don’t want to go to Brazil”
The kidnap mechanic returns.
I really appreciate the effort put in this AI, as a guy who usually don't play online makes my experience really nice
i can already hear Viper SCREAMING "HOW AM I GONNA BEAT 4 AI's NOW"
VHAAAAAAAAAAAT!
"They are even quickwalling? Can you not, please?! Stop microing!"
it will never happen!!!
I watched it like two weeks ago instead of when it was first release, but spirit of the La Hire is haunting me
AHHHH!!!! Why you f#&@ kill another sheep?!
@@TheDemonx2000 "You English fop!!"
The first time I played against the extreme AI (AND WON, YEEHAW!) I was trying to show a friend kinda how to respond to the enemy aggression, quickwall, counter enemy units etc etc. I didn't expect, however, the AI to do a drush. Caught me offguard but managed to survive long enough to beat it. Man, it was so hard but felt so good!
thats what she said. yes it was worth necro replying to this 2 year old post
I wonder if you could do something with having the bespoke AI leveraging a separate machine learning AI.
So instead of having a machine learning AI that plays the game, you instead have a machine learning AI that answers difficult questions that the main AI might run into.
"Is it time to attack now?"
"Should I drop civilian pop to zero to try for an overwhelming push?"
I guess that is probably best left for when the limits of the bespoke AI are being hit and its more or less locked in, and it sounds like he still has a fair amount of room in that regard.
that's actually really clever.
The thing with ML models is that inferencing scales pretty well with model complexity but the initial hit of having that model is substantial. At the end of the day you're better off just doing the whole AI in ML.
As I understood it, the problem is getting the information out of and into the game. There is no interface for a ML AI to command the game. But even if you use ML only for certain decisions, i.e. only using the results of ML, you still need to get all the data out of the game for training.
However... we do have recorded games. So you'd *only* need to convert a recorded game into something usable as input in a neural net. Which probably isn't simple, but actually quite more doable than an interface for the running game.
I think someone could "crack" a machine learning software into the game, but it only could learn about the decisions made at one game. I'm inclined to a different approach: a patch that allow players to be recorded, then send all the data for a match to Microsoft and they could use a system to analyze the data. They don't even need video recording, just record of clicks, coordinates and time. Also, the seed of the map and then feed it into an expert system returning a new AI script.
This could be used as montly patches.
@@MrlspPrt aka recorded game files?
Extreme Ai is challenging but you can kill it, the micro of single units it killer though...
Having its rushes actually look for your villagers rather than attack the first thing it sees would bump it up quite a bit
nice to have this interview
This video looks Promising
How have they not thrown each AI difficulty into the ranked matchmaking to see which rating the AI would end up in?
I think that would be really good. I'm not sure if other players would enjoy playing against an AI when they were trying to play against players, though. As long as they only play a few games per week, and the games are spread out, I suppose people might not complain.
@@Jay_Sullivan I think it would take a little bit for people to realize its an AI. They'd have to give it a unique name like any random player and throw a few out there. Obviously if people know its an AI there are exploits to beat it. it would only work if its a surprise. I'm just curious if like the "hardest" ai is at about the level of the average ranked player or something on the ladder.
@@McSniffle worked my way up from 900 to 1200 the recent months. On Arabia I think the extreme AI would be around 1050, I could actually kind of relate to what it was doing. On other maps not so much though
DAMN IT TOOK LONG FOR THAT CASTLE TO GO DOWN
One. Ram. At. A. Time.
Nice interview! Very interesting
Very clever questions, Spirit!
Nice interview. Just to add to the machine learning, despite it taking millions of rounds for the ai to understand how to win a game, one of the other advantages is that glitches and bugs are sometimes being exploited by the ai and thus being discovered.
I'd like to see more interviews of this format; maybe with the creator of the Barbarian AI, other AOE Devs, or even people in the modding community. Good stuff!
I agree that machine learning for the ai needs access to editing the main code of the game to program it to allow machine learning and develop its own commands and ai commands to evolve and get better. Although one of the things that the ai does very well since the original age of kings was for it to dodge the projectiles coming out of any large projectile making unit: mangonels, bombard cannons, and cannon galleons.
AoK AI did well until you ground attacked, then it got slayed en masse
One feature I'd really like to see would be assigning different difficulty-levels to individual AIs, rather than all AIs being fixed on the same level.
THIS PLEASE! Its so necessary when playing against the AI with your friends who have very different levels of play... Most other games after aok got it.
If the extreme AI would play ranked, which elo would it get? I would like to see it as a 24h twitch stream from the point of the AI :-D
I just hope they can finally fix the villagers bugging out. Every game I find the weird bug where 2 villagers bump into each other and keep doing a little dance for like 30 seconds. Or worse they just freeze.
awesome dev for an awesome game with an awesome community that's been around for over 2 decades now. what a time to live.
Lmao there's a villager that's getting yeeted by an army of skirmishers
If the AI would focus villagers in their rushes, it would be so much more difficult. Whenever they come in Feudal with 15 Archers they attack the first building they see. You can get away with just building no units in Feudal, going straight to Castle Age, and all they do is take down a house and a barracks, before you can engage with knights for example. This is the case also on extreme difficulty.
Me who still plays on moderate HD AI 😎 (thx for the likes)
Tomek T I play on hd easy haha 😂
Yep
@@srpurple4209 Tip, try to build your army early instead of researching techs, you know what a fast castle is?
Still play the hd edition
@@tomekt9675 I can easily beat all AI difficulties on HD but it's so chill vs easy AI so I always fall back to that ^^ Just add extra opponents if the difficulty is too boring. You end up fighting an army worthy of an extreme AI, but with the general ability of an easy going AI.
big respect for the Artifical Intelligence coders they are valuable people
13:45 Bird crashes into edge of map. How have I never noticed this before
Yu are now an RTS deity
I am in awe
Thanks for this Interview, Spirit!
How does the AI avoid getting stuck with something? For example idling units for extended amounts of times, or running back and forth with villagers, constantly switching tasks. Is it all hard work to fix these things one by one, or is there a grand plan to solve these issues?
Wouldn’t it be cool to have an agent of each AI’s difficulty playing on the RM Ladder (even if it were only to play Arabia Random Civ). I’d be super curious to know where they stand
Is there a way to have the a.i. be different difficulties in the same game in Definitive Edition?
When I finally beat the hard ai on de it felt like such an accomplishment. The achievement has only 7.8% achievement rate on steam.
My internet went screwy today so I tried again and beat the extreme ai. I hope I get my achievement when it comes back.
I think the note that the AI will know your army size and act depending on that, will be valuable info for TheSnek. Might influence his playstyle vs the 4 extreme AI.
Wonderful to hear german English. It's so familiar cause I speak in a similar way. xD
I hope Promi will add a scout rush as well. (Quick) walling the entire base sounds awesome too.
My only real problem with the AOE AI is how bad it behaves in long games. I'm not an especially good player, but my cousin and I play a team game every once in a while where we take on multiple or even many AIs at once. However, since we both like the big battles for territory and are pretty vulnerable early game, we like to play with treaties. Or last time we played on the Michi Map so the AI couldn't meet up with us until we were ready for it.
We usually battle the extreme AI because we like its behavior in battles and its fierceness. But even it has so many quirks that always take away something of the fun. And I don't think it's because the AI is bad or something, but just that 90-minute treaty games with 500 population are understandably not what the devs are focussing on, but it would be so nice to change up some minor things to greatly improve the experience for us casuals.
First of all you really notice how it handles villager production and military in a standard 1v1 no-treaty way. It will always produce some amount of military units and stop booming with villagers at a certain point. Which is weird to me, since I've seen it delete units to train different ones before, so I don't think stockpiling villagers is such a big problem for its code. But having 90 minutes of basically free booming, and then wasting so many early ressources on completely useless military units, only to then stop making villagers after 100 or so when it could go up to 500 if it wanted is really weird and makes outbooming it so incredibly easy, that we most often decimate each AI until the last one, and then keep sending ressources to that last player to keep it fighting.
That said, it's also not concerned with having too many useless villagers. Because even if there is absolutely no tree left on the map and nothing to farm, the AI *still* tries to maintain an "economy" of ~100 villagers, even if 80+ of them are just standing around aimlessly. They are not tasked, they don't build, they don't farm, they just rot away and steadily replace every villager that gets killed.
What also happens is the opposite: The map still has plenty ressources, but the AI doesn't renew any camps or farm ressources with more than a couple of villagers. There could be huge piles of gold in its base, but if the game went on for 4 hours+ or so, the AI will just ignore it and have its villagers stand around aimlessly until they die and get replaced (with equally useless vils).
This is really sad because I'm a huge fan of the game, but this kind of behavior makes the games still really boring because it first starts lagging due to all the useless military units running around with nothing to do (but still visiting each corner of the map and spreading out). And once you survive the initial flood of units and destroy part of their village, they either are out of ressources and just become useless, or they don't rebuild and die anyways.
I feel like these issues are not even that hard to fix, they just need to be actually worked on because in the usual competitive 1v1 games, this will never be an issue, thus it will never get changed.
@@Whisper0ak what are you talking about? Did you read what I said?
First of all, to actually answer your question: Yes, I wonder why you'd give up if you had a couple 100k ressources of any kind and 5 teammates to help you out, and you instantly become useless once some basic infrastructure is destroyed. I'm not talking about complete annihilation, I'm saying that the village is broken open and the TCs destroyed if you will. But if you float basically infinite ressources for more military or a complete rebuild of your base, and you have the space and allies to fall back on, what's the problem in doing so? Why even stockpile ressources in the first place then?
But my declared issue was that they don't even need to be attacked or destroyed, but that they just stop acting altogether after enough time has passed. If *they* attack *you* and you destroy their army, and they still have ressources, why wouldn't they train more? If there's ressources on the map, why do you spam an insignificant number of villagers, only to have them be idle for the rest of the game? My whole point was that the AI becomes completely passive over time, no matter my interference.
And considering it's the most difficult AI, it should always keep fighting if there's a reasonable chance to bring it back. Basically surrendering once your first wave of Offense died, with plenty ressources to spare, the extreme AI should not be so demoralized that it gives up.
I seriously don't understand how I'm "being unreasonable" here...
@@Whisper0ak only that the AI hasn't "lost" if it still has the economy, ressources and military to win. That argumentation makes little sense. It's not really close to "modeling real players". Only because you have a very linear definition of how "real players" would play the game.
Ok, but what about AI monk micro.
The way an ML algo could work is by properly generalizing events in game to higher levels and then generalizing the reactions so that the model has less avenues to consider. Chess works because you can break down every possible move into a (admittedly massive) tree of possibilities with each decision and it’s possible payoff calculated. Do the same for AoE and it’s possible.
I think the question on all our minds is... when will the AOE2 AI become self aware???
should have asked why eve the extreme ai still keeps attacking random buildings until they go down
It must be fun seeing this stuff beat Viper in a 1v4 and he's amazing at everything!
Good work with Promi and I love your videos Spirt!
I have watched every one!
What I'd like is some kind of exposure of functions for a c-styled language to make AI with.
Seconded! No clue if it is possible.
Respect to you Promi
When a game is played . It saves file . It can replay. We can select opposition team. What did AI did?
How this file saves each players single activity. Please explain.
The blue mango at 7:50 just yeeted his shot in a crazy way
Edit: Same with the red mango at 11:59
@Emre Vural maybe true, you remember that flying mangonel on T90's channel, maybe siege Is busted
I really like this type of content....gives some good insights...👍
Spirit of the Law please make a video about building. Things like, how much hp do buildings have as soon as you place them? how many resources do you get back if you delete them before finishing (at different stages, if the building was damaged)?, how do using multiple villager speed up the build time? if I want to produce skirms urgently should I build two archery ranges with one vill each, or one first and then the other with both vills? etc.
One question that wasn't answered, is the AI capable of playing properly at different maximum population? For exemple, is the AI balanced for 500 population or is it made for 200 pop?
Tested multiple times - seems like it's made for 200pop as vill count is always the same regardless of the pop size.
With or without machine learning, the hard AI beats me like its no man's business. I need to be scripted, not the AI :D
Also, I would love to see Viper's reaction to the latest changes in the patch that include the improved pathing.
How do you get such a high quality audio? Did you record separate tracks for you and the programmer?
Would AOE be the most difficult game for an AI like alpha star to learn ?
I can't think of an other game that has so many parameters, randomness, where macro is so much more important than micro, 30+ civs to learn, so many different types of maps to learn ...
I think it won't be that hard given how good it has gotten on starcraft
@@lalramdinavarte4987 Starcraft has more emphasis on micro and less on macro than AoE. Starcraft has less map generations. Starcraft has less civs. Overall, there is probably less decisionmaking, which is always the really hard part for an AI. Pure micro with one strat can bring you really high on StarCraft, it will never work in AoE.
Exciting to see if there's going to be one too, the progress will be entertaining to watch.
SOTL did mention long-sword man when It comes to the Alpha-aoe2, so I guess It's not gonna be good under the condition.
Look up Dota 2 OpenAI. Much more complex problem to solve and also requires teams as there's no real 1v1.
@@R3DT1D3 Definitely not. A MOBA is relatively straightforward and the fact there are 5 heroes to control does not change much for an AI. The real difficulty is to add a proper playstyle for each of the 100+ heroes, but laning is quite easy to understand for an AI and even complex lategame strat are a lot less complex than in an RTS. Simply the fact that you'll almost always walk in a straight line towards the enemy base or occasionally flank is a lot less complex than in AoE where you need to set a priority for each area/resource collection, to know where to castle, where to hit with archers, how to loop with your units to hit the enemy base without being seen, where to protect with spears when the enemy will scrush... And lategame when and where to raid with cav...
I think what would help to make it more human like is to possibly limit its APM and increase it per difficulty level, because I remember when TheViper was trying to defeat the AI, but had extreme levels of micro that were far superior than any pro level human. Also I think giving the AI some more human tactics would help as well. MAYBE even some meme strats lol
If you allow pausing you can get on the level of the AI for micro.
15:46 this is why i never use mangonels
Hey SOTL, I would like to see a video about the comparison of cuman 2 TC feudal boom vs 3 TC castle age boom!
So my friend and I recently started playing the game, and I still have some questions about the options within the online multiplayer lobby. The options are AI, AI (CD), and AI (HD). The HD clearly is the most difficult to us, and a moderate HD tends to play about the same level as the Hard AI. Is the HD one receiving the live updates discussed in this video? It doesn’t sound like it from the way he talks about HD here. Can someone give some clarity as to which setting is getting the live updates, and what the other options were from? (AI/CD/HD)
I assume you're playing the definitive edition of AoE2. The "AI" (often called the DE AI) option is getting updates, not HD or CD. It's interesting that you say that the HD AI is harder than the DE AI. I've found the opposite to be true.
Hi! So, the three AI options are from the three versions of the game: The original (AI CD), the HD edition (AI HD) and the new DE edition (AI). ONLY the DE edition is getting live updates. The other options are there only for historical completeness (so people nostalgic for old versions can continue to play with the AI they are familiar with). The 'skill levels' of the AI editions are not directly comparable - for example, the HD version has some weird difficulty jumps which were smoothed out for the DE AI, and DE AI got an additional setting. However, if you put a DE and an HD edition AI on max difficulty and let them slog it out, you'll find the AI typically beats the AI (HD) pretty easily.
I know this is a niche thing to ask for, but could you do a video on water counters? Like what ships counter what, when it'd be better to use your land units to deal with water units and vice versa, etc
Lovely and intresting Interview. 😊
I always wonderd why it isn't possible to selectthe AI strength for ever, AI player in the game individuell. It would be really great for balancing if the human players aren't on the same level. 🤔
For ML AI I think I can do it. What you'd need is something that updates game states regularly
Regarding machine learning: I think it would be reasonable to implement ML for small tasks (individual combats maybe?) while leaving most of the AI to for/while/if/else statements.
The ai could be much stronger if they have that, sadly there is no one interested in a project of this type. And if is posible to add few ML behaviours and not broke the ai, its obvius that a full ml agent would be much stronger but would also take more time. Maybe Aoe2 Alphastar would take hundred of times more time than the Sc2 alphastar. At this point i dont know if it would be worth the efforts.
Promi, the only person who get’s to beat TheViper over and over again. But Promi, thanks a lot! I usually play against the ai and it is a lot of fun!
A couple of questions from my side,
does the current DE AI (including Extreme AI) has the ability to strategically onager cut forest / trees to make way to enemy town?
Secondly, does current DE AI know how to use attack ground on mangonals for micro?
If answer is no, is it planned in future? Thanks
Great interview! Any chance of interviewing a former Ensemble dev?
Given the Conquers AI ran on a 233mh Pentium, and the new AI takes far more computing speed and power, the original AI's ability given the limited resources is quite remarkable.
Has anybody a link for Promis´ Twitch channal?
It is in the description
Technically knowing the location of an enemy is not cheating as long as the location was explored. Constantly querying the map for buildable areas, you can guess where the enemy is. It is like trying to build a castle on a hill but the silhouette comes up red so an enemy is already on the hill.
Love your content Spirit, can we see AI using rams to take arrow fire...?😎
8:30 I think the step between moderate and hard is now big. I can beat moderate AI pretty comfortably on any map with any civilization but get my ass kicked by hard AI everyrime :-/
Yeah same, I think now to beat the Hard AI I am gonna have to git gud on Ranked... That's the point i'm at.
Machine Learning, a fascinating things to play with. I don't thinks its because the ''map changes everytime'', its because of number of 'model based reagents' are much higher than starcraft or DOTA2, a particular model handling takes 10 to 11 neural networks on DoTA and all AI has to do is control and command one character. In AoE its build, maintain, gather, and attack! 4 model based reagents are difficult to control on current PC configurations. So evn if they build it, it'll be only playable on supercomputer facilities.
yeah, I also think it's not, in reinforce learning the algorithm designed should be capable to detect environment dynamically already (which is already a difficult part), and an AOE game itself can divide into many parts more than you mentioned I think, and overhead of that NN cannot fit in PC and more importantly is extremely costly for both building that up and running it
I've noticed it almost never ever scout rushes. I'm guessing it's due to the difficulty of unit micro.
I would say the AI knowing how many units you have so it knows when to attack is definitely a cheat. Also based on the video the AI is still suicidal, throwing units away to TC's, archers ect. That was HD AI's biggest weakness because you could just wall and castle up and they'd just suicide everything they had and eventually would just stop doing anything (I don't know if that last part is still the case).
I feel there is too big of a jump in difficulty between the Hard and Hardest AIs. Any ideas why?
I can easily beat the Hard AI 100% of the time. But I've been yet to beat the Hardest even once.
Just keep playing, you will have to keep at it until you beat. That's how I did .
I find that hard ai is quite slow to switch to counter your army, while hardest ai is very quick so you either anticipate his counter, or push hard to win the game before he can switch. Just my experience
The ai now is definitely better than it's ever been. That being said, I would love to see it act more like a real player at difficulties other than extreme. I like to play the ai on hardest when I'm practicing strategies for online because it's similar difficulty to paint against players my elo (~1000). It would be cool to see the ai on hard/hardest attempting advanced tactics/strategies without flawless execution. For example making scouts and making 1 or 2 too many farms then struggling to transition to archers later. Or occasionally trying a scout rush, drush, fc knights, etc. For me on open maps the only strategy the ai does is fast feudal with massive amounts of archers and skirms. At my level I can put produce them but they're also not smart so I just fc knights and win. They're really good at eco management, but not realistic about using civ bonuses, military strategies, etc. Would be cool to see the ai play like a player at my elo: attempting advanced strategies/good counters but imperfect execution.
When will the AI be capable of more aesthetic farm placement?
okay but why does the AI at lower levels just implode on any water map, especially migration-style maps?
Thanks for the video, but is there any documentation for the new AI features?
Hey Spirit, awesome vid again. You should do an AoE 2 de Olympics! That would be a great, GREAT video. Greetings!
Spirit, have you ever done an 8 player match on a tiny Arena map? Quite an experience.
Where the cumans overview at?
It's weird because I usually play Standard or Moderate AI's and the HD seems to be much more aggressive and generally harder to handle than the DE one, although the HD seems to completely overlook building castles and mining stone in general.. Does that mean it gets harder as you go to hard/extreme level? I have a hard time finding the balance between getting anihilated and destroying ennemis super easily..
Moderate to Hard in DE is pretty big jump but still doable. I prefer the DE AI.
Question: Do you reckon there will be a cap to the expert difficulty and if so ever plan on making a quote on quote impossible to beat Ai
"And I'll see you next time".
Wait... we can't see him but he can see us? I believe he's really a spirit! 😱
I guess what is needed to get a machine learning AI is some API to access a running game from the outside (similarly to what a human player can do), instead of a script running inside it. Then you can use whatever generic ML technology you want to play and improve your game.
In response to which AI difficulty step up seems to be the steepest - I feel like it's the Moderate to Hard jump. I've only played Single Player, because I can kind of keep up a build in the first 30 villages, but then things get really overwhelming to keep track of, really fast. I can crush the moderate AI, but then once the Hard AI gets going, I find I have no clue how the AI has a 50 unit army and comes and rushes my base, and I've got maybe 10 or 20 knights.
Does anyone know if there is publicly accessible source code for this? I would love to take a look at how it's being developed.
No they won't give
OMG i already loved TheMax but know that I realize he wrote Barbarian i reeeally admire him.
SOTL is a great interviewer...
Didn't anyone see the Terminator films??? He's working towards Skynet dammit!!
I love SotLs enthusiasm for Machine Learning AIs but it IS crazy complex and would take a lot more than just Promi working on it. Just making an AI that FUNCTIONS in any shape or form is horribly difficult.
this guy is doing god's work. this game will last forever because of his AI. I mean, HD's was already cool enough. This is just going above and beyond.
i wish blizzard games got as much love for their AIs.
What about the creator of the barbarian ai. Themax ? His ai’s was sooo competitive compared to promi
pls do cumans overview pls
Thank you!
Which AI is harder to play against, the moderate HD AI or the moderate DE AI?
The hd
i would like the extreme AI to wall up more to secure late game eco