Biggest possible improvement to the AI would be to make it handle its army in organized grouped way, like humans, instead of current individual micromanagement. This would be awesome, but probably very complicated to develop.
Agreed. Also they could develop this for the AI :
-
Kiting with archers
-
Ignoring random buildings when attacking, and going straight for ennemy eco and/or ennemy army, while keeping out of range of town centers and castles.
Only attacks buildings/walls if there is no other way to reach opponent base. -
Switching targets of attacks, and not just go tunnel vision from its base to the other player’s base.
Attacking from sides for example, if there is a wall or a Castle in front. -
Mid-Late game raids with the appropriate units (knights, eagles, light cav, …)
-
Wall his base and/or place defensive towers or castles if needed.
Yeah, it’s complex and takes time to code/develop. Devs might probably never do this because it’s probably not worth the time and ressources invested. However, it’s possible. Some other RTS have designed AIs capable of such feats. (AlphaStar anyone?)
1 Like
The construction of the AI is a completely other level then the way an AI is constructed at AoE II. I even dont know if it would be possible to implement such thing into AoE II.
AlphaStar is a trained AI based on neural networks. Based on data the AI preformed better and better over time, as it gets more and more data. The AI in AoE II is just a set of action what the AI can do based on some scripted statements. It just dont learn anything.
I really do like it if someone can implement such AI, but I dont have the knowledge to build such AI. I even dont know if it is even possible for AoE II to build such AI.
If someone is able to build a strong AI, it is TheMax. He is a pro player, but also the maker of Barbarian, the best AI before DE. I hope he will update his AI for DE. That would be a really strong AI. I wont even try to beat that AI on Extreme. I will just loose without doubt.
Pic 1 shows the time I start attacking his wall:
Pic 2 shows when I break through his wall and start building a barracks. Note that the villager’s loss of HP is from a wolf on the way over, not the AI attacking him.
Pic 3 I have started building an archery range, still not been attacked.
Pic 4 I have displaced him from his gold mine with only my scout, have a few MaA to harass him with. I have built a mining camp as I realise I have no gold! (microing inside his base really hurt my ability to function back at my own base).
Pic 5 - he has got a castle up, and it has taken out my scout, but it is attacking my gold mine rather than my villager. He doesn’t have murder holes, so my MaA can just stand there attacking it.
Pic 6 - I have a small number of archers around that are crippling him, just like worked on Arabia. The castle has taken out my mining camp, so my villager has retreated. I’m still able to deny him using that gold by having an archer on the other side of it. You can see I have an archer on the other side walking past his TC. I’m keen to get one into the little spot by the trees out of range of both the TC and castle.
Pic 7 - you can see the disparity in response to a tower vs archers. He sees my tower just outside his walls as a huge threat, and sends many villagers to attack me, successfully preventing me from getting it up. But even though an archer set to stand ground is nearly as effective as a tower, he never sends multiple villagers to rush at them to take them out. My one archer over on the left is, I think, 1 square out of range of the TC’s fire, and is denying the use of that gold by the TC and crippling a huge chunk of his eco. You can tell I’m doing huge damage, as my score has nearly caught up.
Pic 8 - I have sent a couple of crossbows to deny the gold at the top right of the screen. My 2nd TC is going up. I have a crossbow inside his base in the top corner, above his 2nd TC. I’m able to get them there because he doesn’t have ballistics, so they can run straight past the castle. My score is now 50% higher than his.
Last pic, you can see it’s all over now:
So, the main things that would have defeated me are:
- Respond to me attacking the wall in some way. Could have built something inside that point, or sent 3 villagers round to attack the scout and villager.
- Respond to me getting inside in some way. Could still have sent just 3 villagers and won that fight.
- Respond to lone enemy units with villager numbers. Could have killed my scout that was denying use of the gold mining camp, and could have killed the lone villager that built all the buildings inside his base. Could have responded to the lone archers like it responded to the tower going up.
- If his castle had attacked my villager in preference to the mine, it would have died from a single volley, and that would have hurt me quite a bit.
I think most of those would be very easy changes to make to the AI. From previous patch notes, I think it already has the ability to predict the outcome of a battle, so it should be able to work out how many villagers are needed to successfully rush at a threat. Making the castle prioritise villagers over a mining camp would be easy, and I think would always be the right thing to do. The only thing that I can see being tricky is making it build something inside the bit of wall being attacked, as that requires significant understanding of spatial relationships. It probably is still possible using a small evolution of the pathing logic, as it must know that there would be no way for me to get from the outside to the inside, and it must be able to do a “what if” on the scenario of that bit of wall disappearing, and know that it would change the situation and allow me to get through when I couldn’t before. So you then just have the question of which adjacent position to build a house, and it could do an exhaustive search through the few possible adjacent positions until it either finds one that prevents passing through after it has been built and the wall section is gone, or has tried all possible adjacent positions without finding one that achieves that.
1 Like
I did say that because i was amazed about it worked. I didnt say it because i didnt believe you.
I think this this is really an exploit to win against the AI on Arena. Even against 700 elo players, you wont get in with just one vill and one scout. If you really want to know the elo of the AI, then you dont really need exploit strats.
I even think my krepost rush is a doubtful case. I dont think it will work against most players around 1000 elo. Maybe i still win, but the players will defend better. For example: The AI did build castle. It was in a pretty much useless position. It really didnt anything for the AI. Except that i just didnt push that way, but in the end the AI also had nothing in that area of his base.
I think it’s valid to try to find weaknesses in the AI and exploit them. I’m sure the top humans, who often play against the same humans many times, will do exactly the same thing, and try to find weaknesses in each specific opponent. And then their opponents will adapt their play to try to eliminate those weaknesses, and similarly, the devs can use the information I’ve posted above to try to eliminate these weaknesses in the AI.
The AI could similarly exploit human-specific weaknesses. The AI is very good at microing all over the map at the same time, but humans find this much harder. So an “exploit” the AI can use against humans is to always try to overwhelm the human’s micro by creating action in lots of different places at the same time. This might not make it appear better when playing against AI that doesn’t do this, but it should improve its performance against humans.
That is only valid for the pros. At the elo of the AI, you dont play frequently against the same person. So you cant really exploit the weaknesses you already know. So if you would meet the AI in a ranked game, you dont know it will be the AI and you dont know you can use these weaknesses.
I think we both agree: Against no real human you can break the walls with just 1 vills and a scout and than build the military buildings inside his walls, without him doing anything against it. So without knowing it is the AI, you would never go for this strat. Therefore this wont really be a strat to test the elo level of the AI for me. If it is just about having fun: Do what you like. Exploit the AI completely. I really dont care. But it wont give the answer to the question: What is the Elo of the AI?
In theory the AI has much higher APM than humans and thus can do a lot of micro. In practise is dont see this happen. Based on my current games against, i have to say the AI is really bad at microing.
What it has done, though, is helped me to understand the comments from players in the threads I linked to in my opening post. I thought I must be unbelievably bad because people were saying it’s so easy to beat the Extreme AI, but I was finding it quite tough against even Moderate AI. But I see now that the people who are saying it’s easy are almost certainly doing it in a similar way to what I’ve now learned to do, e.g. using archers to camp at resources. So comments like
“I can beat Extreme AI but still lose to Hard AI if they mass a huge Siege army which kind of shows that the AI can reach another level late game.”
“If it gets to late imp and can spam units it’s nearly impossible to beat. And if you play cheesy and tower rush or do an archer rush where you camp it’s resources, it’s easy to beat it in 20-25 min game time.”
make sense to me now.
So there are really two different questions:
- What is the lowest ELO human player that can learn how to beat the Extreme AI by exploiting its weaknesses?
- What ELO would the Extreme AI be if it played in ranked 1v1 matches without people knowing they were playing against AI?
Now I know that I can beat it with those strategies, I can move on to trying other strategies, e.g. I’d like to become able to get close to TheViper’s Turks Fast Imp time on Arena and beat the AI with that. And it would also be fun to play against the AI where it’s doing Turks Fast Imp, but build a defensive counter-army and see if I can do that fast enough to wipe out his first wave and win from there.
If anyone wants a real challenge against the AI, you really need to wait till its first attack after it has reached Imperial. Until that point, all you’re allowed to do is build and create units at your own base. I think after that point, the “cheese” strategies won’t work, because he’ll have military that will swiftly attack any units you try to camp at his resources with.
2 Likes
I did a ranked 1v1 game, and was matched against a player with <800 ELO. I lost, and they appear to consistently be around that standard, so it seems even a <800 ELO player can sometimes beat the Extreme AI. Even with the “archers as towers” strategy, I don’t win every time, it’s finely balanced as to whether I can identify the right places and get them there early enough. I think, though, that the player I played against would probably lose every time to the Extreme AI if they played the way they did against me, as they didn’t apply any early military pressure, so the AI would progress to its strong period.
Not applying early pressure is one of the big weaknesses of lower rated players. They tend to play to passive. Many of them will have a higher rating if they will apply early pressure.
The DE AI is still really weak compared to a human player. You can consistently beat it with a ELO superior to 1200 which is just average. Just for comparison, the Barbarian AI on Hard, that is the highest difficulty for it without relaying on cheating, has an ELO of 1350-1400. Also is the better one to train the player on a real multiplayer scenario on an open land map. If you want to get better i recommend practicing with this one.
I like to practice some things against extreme ai. I’m a 1200 elo player and it’s fair to say the ai is extremely bad at some things and extremely good at others.
If you rush it, it’s terrible. Still bad at tower rushes but the recent patch improved this greatly as if it sees an obvious tower it will defend with vills. You need to be more careful about chaining them well now and getting one up out of sight before a closer one. Scout rushes do little damage to extreme ai also now with micro and defending with vills. You can pick one or two vills off but it macros well enough for that not to make too much of a difference at this level.
Overall it seems fair to say it might average a 1100 player but then it’d be like a 800 player for being rushed, certain build orders, and especially map control (versus towers/castles and placing its own) and a 1400 player for certain things - once macro gets going, countering units, expanding on map, and especially for scouting.
So it’s such a wide spread of good and bad points, which a human player is unlikely to have at this point. If you’re that good at micro, your map control and defending rushes won’t be that bad for example. And likely similar spread of skills for the lower ai ranks.
It’d be great to have ai in the matchmaking system and see how updates improve it. I think many people would enjoy seeing devs posting that extreme ai is roughly 1100 elo, for example, or wherever it ends up. And then it’s good practice in many ways. You know if you don’t cheese the ai, then picking certain ai will give you that challenge for trying something or practicing something, etc. Making sure you’re walled up in feudal while fast castle or after drush. And for many players it’s the confidence to start online games. When they start levelling up against ai.
I’d be very curious to see the average elo ratings of each ai 
hope the devs do it
2 Likes
To really test the AI i went 1v2. I picked Arena, so the AI have some defenses and i cant really rush the AI down. On the other side i didnt go for extreme immediately. I settled on Hard. I think we can all agree that 2x Hard AI is more difficult than 1x extreme.
It was still kinda yolo. No real focus, something i really need against humans. This was just building krepost until both AI where destroyed at 46:28. It no stage into the game i was thinking ‘i can loose’. One AI hit imperial age, but it didnt matter. After the game i saw that one of the AI had much more resources gathered. I really dont know what he did with all resources. He beat me on Food, Wood en Gold with ease.
The biggest weakness to the AI seems really making army. After 46 minutes the biggest army of the AI during the game was 12. The other AI largest army was 3. That is pretty terrible too me. I must admit my largest army was 12 too, but i had krepost killing / destroying all.
Almost krepost only: 376 kills, 47 units lost. Most of the losses where Konniks without any upgrade… I also destroyed 36 buildings…
Based on this game, i feel thing i can take on the extreme AI, in 2v1, maybe it will be really challeging. Biggest challenge was where to build the krepost. The AI likes to really build everything in a small space.
1v2 - Take 2
Again Arena, this time on hardest. Again i went for krepost rush. This time i even added some towers. The result?
521 units killed, 50 units lost and 52 buildings razed. The game ended after 55:55 minutes. Largest army of Mayan AI was 7, largest army of Malay AI was 17. Not really impressive. The vill count of the Malay was impressive: 133. Even while being pressured, he managed to get 133 vills.
This AI can be really bad sometimes. I beat both AI by over 4 minutes (!) to the castle age. Note: One of the AIs was even Malay and thus advanced more quickly! That is just really really bad play by the AI. One of the AIs even failed to hit castle age before 20 minute mark. No joke. I even think 600 elo players can do a 20 minutes FC…
I looked at some Arena games, looking at scores at the 40 minute mark at a few different ELOs:
750 ELO 6220
950 ELO 7550
Extreme AI 9600
2300 ELO 11580
So I plotted those in Excel to see what ELO would fit with the other points on a curve, and it’s around 1500:
Obviously that’s only one element of play, and we know from the discussion already that it can be beaten by play of a much lower standard if you don’t let it operate freely up to that point. But I think it shows why TheViper hasn’t been able to beat 4 of them yet. He isn’t really able to hurt more than 2 of them early on, so there’s always 2 that are able to play like that, reaching a combined score and ability to spam military that far exceeds his.
1 Like
Tatoh the pro player spanish say that the in the moment tha resources and codes do that be impossible made work in on alphastar in the system, but I dont know if is possible
I really dont think score is a good indicator. I had a look at three games on Arena, at 40 minutes. If i have to believe your list, my elo would be around 750, but it is around 1100. Also i have never seen extreme AI around 9600 in my games. So i wonder which kind of games you used.
Can someone tell me why the AI is programmed to to the following builds on Arena:
- 27+11 FC
- 33+7 FC
Difficulty was hardest, not extreme…
9600 was Extreme AI vs Extreme AI, Vikings vs Tatars on Arena. It did a slightly odd FC build, in that it went to Feudal with 23 villagers, reaching Feudal at 10:56. Castle Age was reached at 15:58 with 28 villagers. From reaching Castle it went straight into a 3 TC boom, and it reached 100 villagers at 28 minutes then advanced to Imperial.
TheViper recently raided https://www.twitch.tv/promiskuitiv , who develops the DE AI, and he was good enough to spend several hours answering questions about the game.
One thing he mentioned is that the AI does still “cheat” in some small ways, in that it uses information that wouldn’t be available to a human player. One of these pieces of information is the total amount of military that the opponent has. So the AI knows how much military you have, even if you hide it at the back of your walled base to prevent it from being scouted. I think this explains some of the things I’ve seen:
-
When I send a small number of archers against the extreme AI, it will fight them with villagers, but when I have more archers, it stops fighting them, even if there is still only a single archer attacking them. I think this must be because it is regarding them as being up against the total army, not just the one archer that is close to them. So this is why you can dot archers around, one per resource, and completely shut the AI down.
-
I think it also explains why tower rushes are so effective, because you haven’t created an army. I’ve found that if I try to mass numbers of military at my base, then when I go to the AI’s base, it has greater numbers, and I think this is because it’s responding to knowing how many military I have made. It always knows the size of my army before I get there. But with tower rushes, you have no army, so it doesn’t see it as necessary to make its own army to counter yours.
If i hear you talking about the game, i just feel like i play a different game. For me the AI just dont build military at all. It never had greater numbers then me. I got game where i had 20 military units in his base, but he has almost nothing.