AI Development Discussion

classic Classic list List threaded Threaded
1250 messages Options
1 ... 34353637383940 ... 63
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

something interesting
I played two games on the Middle Earth map against the AI, here are the saves if you are interested. I played both until I felt victory was clear. I was expecting the AI to do well, since the map is heavily land combat. However, the AI's big problem was not coordinating the different powers or expecting the opponent to coordinate, which has been noted elsewhere, as well as some odd naval moves.

LOTRAIgame_versusgood.tsvg
The AI plays evil, I play as good. The AI goblins play pretty well, others not so much. See the above for why, coordination and strafing are necessary to beat Gondor and hit and run works well for Saruman generally.

LOTRAIgame_versusgood.tsvg
I play evil, AI plays good. AI plays much better in most areas. Freefolk is played really well, except it buys way too many hobbits. Elves also buy too many hobbits. Dale shouldn't buy those rafts, otherwise it is decent. Dwarves are played pretty well. Splitting up forces to different fronts really hurt them. Gondor tries to be offensive against Isengard, when it should be defending against Mordor and friends.


Some of this is map specific, or already pointed out, but what I was wondering about was -
- how does the AI handle strafing attacks? Will it attack expecting to retreat after doing damage for 1 or 2 rounds? It appears it does not anticipate the opponent using strafing tactics. In some of the more complex maps, including this one, strafing is very important.On the common WW2vX maps the threat of strafing a major stack alone can be important.

- does the AI adjust for low luck? Low luck also can impact the game in a big way, especially strafing or use of two-hit units, since one can be certain of the maximum hits possible to give/ receive.
(example- attacking three spearmen with one nazgul is a fine move under LL but very risky with dice)

Those two features would be relevant for every map. Hope these saves can help you
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

redrum
Administrator
@something interesting - Both saves are the same (AI as good) :)

The AI doesn't currently coordinate attacks or strafe territories. These are both on the list of things to implement but are rather complex. These are 2 concepts that are even difficult to teach human players as most people think about whether they can a win battles and 1 turn at a time.

The AI does consider LL in some aspects and since it uses the battle calculator it automatically picks it up when simulating battles.
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

something interesting
@redrum - Here is the other save - ai_game_good.tsvg

Do note that the AI does strafe sometimes, just not on purpose. Strafing can definitely be a one turn consideration just for a short run TUV gain. I know the current code seems to prioritize taking territories, but I would argue that is a bad trait many newer players have. It handles TUV exchanging really well and strafing is an important extension of that.
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

Black Elk
Some strafe logic for the TUV trade would be cool. So far I haven't seen anything like a solid strafe, just a retreat when the defender puts up too many hits and diminishes the attacker's power too much in a single round of the combat phase. Here is a game where that happend to The HardAI Russian attacker. In this case they bounced out of Ukraine to save their fighters, after the first round of combat failed spectacularly for them.

It was a decent start of HardAI Allies too. Gave them 18 ipcs, split 12 to Russia 6 to UK. Which the Russians used for tanks, and the Brits used for inf.

In this case their failed attack on Ukraine was pretty major, even with some bad trading in Air for Axis with Germany and Japan losing a number of fighters, it still wasn't enough to overcome the major set back for AI Stalin in the first round.

HardAI_Allies_vs_Elk_Russia_Ukraine_fail_round_7.tsvg

Here G takes Moscow in the 7th
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

redrum
Administrator
@all - Just committed latest changes and updated the pre-release:
- Lots of bug fixes and balance enhancements
- Improved naval vs air purchase consideration
- Added support for V1 rules of unlimited production in starting factories and limit of 1 AA per territory

Please test and let me know if there are any issues.
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

eurofabio
In reply to this post by Black Elk
What is this strafe attack??
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

redrum
Administrator
The Strafe Attack is defined as "an attack on a territory that is designed to weaken your enemy's front lines in some manner, in order to leverage a tactical advantage for your forces, conducting the rounds of combat necessary to sufficiently deplete your opponent's units from the attacked territory - then retreating those attack forces to safety before you would occupy the attacked territory and before you lose any expensive-to-replace attack units of your own in the exchange."

http://donsessays.freeservers.com/deadzone.htm
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

Cernel
Strafing is the biggest gamechanger with LL versus dice.

Reason being that you are not allowed retreat if you killed all enemy units. So, with Low Luck, you can quite easily make a strafe being sure to leave 1, or almost 1, defending units, and retreat; while with dice strafing have to be part of a strategy taking into account the possibility of taking the territory or not, depending on the dice.

An example being the mass attack and retreat in Athens round 1, when playing Greek in 270BC; which is much easyer to do with LL than dice.

But strafing is not only for TUV destruction. It also has important movement implications.

Sometimes you even make a strafe with an unfavourable TUV swing, if it allows you to have a good move.

A subset of the strafe is the so-called, "teleport"; when you attack and retreat with the only aim of increasing your actual movement towards some other territory.

An example, in round 1 270BC, if Egypt doesn't take Damascus (because it strafed it), being attacking with 1 unit from Damascus and a bunch of units from Seleucia, and retreating in Damascus; so to teleport your units in Seleucia right in Damascus on round 1.

This is particularly stupid because, this way, enemy territories allow you to move faster than friendly ones (another good reason for having greater movement during non combat than combat move).

For example, if Palmyra would be friendly and passable, you would be unable to move your movement 1 forces through it in 1 round from Seleucia to Damascus.

So, if in a game there are important strafes on round 1, it is easy that the balance in dice and LL of the same game is significantly different.

I personally think that it is stupid that you can retreat only if there are enemy units left and, instead, if you killed all of them, you just attacked to well and can't retreat; so I suggested adding a property for allowing choosing to retreat or consolidate after having killed all enemy units.

But that's developer stuff... And most likely such a thing might be added only if an actual boardgame adopt such a rule.

Regarding the AI, that's something thricky to do, and the AI would really need to discriminate between LL or Dice (or just considering doing strafe only if it is LL; since not doing strafe in Dice is really not a big deal).

A warning would be not to end up like Dynamix; that had such advanced features, but ended up being actually worse than even the first versions of Hard AI. Trying to make the AI doing advanced things has the risk of making it worse.

Also, since, as long as you play Dice, strafing is not very important (cause you can't retreat if you kill everything), I think that there's not really a need of having the AI thinking about strafe, because it would make the Dice play only modestly better off, at most (I think it would be a thing to add only when the AI is really playing well).

For example, coordinating attacks is surely much more inportant, in most maps, than strafing.
History plays dice
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

redrum
Administrator
@Cernel - Agree with most of your points. I definitely think coordinating attacks and defending against coordinated attacks are more important than strafing. Another point is that when you have multi hit point units that heal every round it makes strafing more common and powerful.

The 'teleport' mechanic always makes me laugh. I don't think I've ever used it in a game but its definitely strange to think about.

Adding advanced features is definitely a slow process because of exactly what you describe. Need to make sure it doesn't make the AI worse and that it is only using them when appropriate.
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

panguitch
Speaking of multiple hit point units: does the AI understand whether a unit requires another unit to be repaired, like when a battleship needs to be adjacent to a harbor or something?

Greyhawk Wars
TripleA in the original Dungeons & Dragons world

Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

panguitch
Testing the latest pre-release and wondering if something was changed or reverted with how the AI protects mobile factories.

Here are three screenies. One before non-combat move, one after, and then one after placement of new units. The AI blatantly leaves two factories unguarded.

beforeNCmove.jpg
afterNCmove.jpg
afterPlace.jpg

(The culprit in question is blue/Furyondy. If you'd rather see a save game I can upload it, but you'd also have to install the sandbox version of Greyhawk.)

Before this pre-release version the AI never would have done this. Since you first told it to protect factories, it's been very careful. Which is why this caught my eye.

Greyhawk Wars
TripleA in the original Dungeons & Dragons world

Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

redrum
Administrator
@panguitch - No the AI doesn't understand any repair requirements currently.

There shouldn't have been any changes to mobile factories. Can you upload a save? Does it occur on other maps?
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

panguitch
AI understanding repair requirements would be great, but I understand it's a minor thing for a minority of maps.

Regarding the unprotected factories: here's the save game for the first example: unprotected_factories.tsvg
To open you'll need the sandbox version of Greyhawk: https://www.sendspace.com/file/2j8tn1

To try and save you the trouble of downloading the sandbox version, I tried to recreate the situation using edit mode in the current version of Greyhawk (0.9.9). Not surprisingly, it didn't behave exactly the same. And the first time I ran it, the AI didn't make the mistake. But the second time I ran it, the AI made pretty much the same mistake, leaving three unguarded factories: replicate2.tsvg

Out of curiosity, I just tried it using what I believe is the immediately previous jar. And the AI still did it. Save game: replicate3.tsvg

So perhaps it's nothing to do with the latest pre-release. I wonder if the AI thinks the fortification in Crockport is protecting the factory and/or preventing blitz. (Really, the fort, just like the factory, is infrastructure that's destroyed when captured.)

That might explain it. Really, Hard AI has done marvelously in protecting its factories, up until this particular situation. I'll keep an eye out to see if it happens in other situations.

Greyhawk Wars
TripleA in the original Dungeons & Dragons world

Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

captaincrunch
This post was updated on .
Wow. Congratulations! This AI.jar has turned a corner and is now even way more challenging!!! At least a few times I was saying wow this is hard lol.



I had to play 7 matches to get 2 wins!!!




Thats a record amount of matches needed to play and I can't remember every match but I'll try to remember anything significant.




@Something Interesting, Black Elk, Cernel, and Redrum about Strafing Attacks;



My fav "Strafe Attack" occurs when I get to "Teleport" and move some units when you know you will be retreating and not finishing a battle just to do some damage ... the best example is having your units stuck in Finland/Norway and you attack Karelia but retreat half way through the battle and then you can choose to retreat your units to East Europe and reunite them with the other half of your stack of units and I do this all the time when I don't want to leave my low number of units after a heavy loss battle.




@Redrum;



Match 1-2: I noticed right away that the AI.jar plays Russia much harder and different right now. I literally almost don't even know how to open up against it when I am playing Germany. Congrats and I'll try to remember any significant details. I lost my 1st match against this AI where I forfeited my 1st turn with Russia for fun (I was the Allies) and I almost won that match but I forgot to keep an eye on the Japanese AI in West USA and they landed and held West USA and that ruined my best match ever that I didn't but almost won. I then surprisingly beat the AI where I was the Allies and I didn't forfeit my 1st turn with Russia for fun and did not have any problems - I took Germany in round 5 and then took Japan in round 12;

latestaxisaidefeated16.tsvg






Match 3-7: I then went on to lose 4 straight matches until finally winning against the Allies AI! I had a couple matches where I didn't play well like forgot to attack something or sent in the wrong units or didn't play well but many times I was saying wow this is a hard AI now. The match I finally won I didn't even have any plan because I was getting so pissed off at losing and I don't even know how I won it so easy but somehow my little victories with Japan helped and somehow I held off the Allies AI Navy landing into Europe and won. Redrum, I took the Russian AI capitol in Round 4 but I want you to see how the AI defended the capitol ... it basically DIDN'T. I kinda chuckled. Check out the game save at round 4 (the AI left just 1 UK Plane there but I had a match I didn't save where it left 0 units in the capitol and I had a clean Blitz Tank for Japan with an open Territory to take the Russia AI capitol for free which I did). I then took the UK AI capitol in round 17 and then finished by taking the USA AI capitol in round 20 (latestalliesaidefeated16 HARDEST WIN EVER);


latestalliesaidefeated16.tsvg





So, congrats on an officially challenging AI. I noticed the AI attacked Hawaii with Japan in the 1st round again in 1 match I lost ... but then the AI DIDN'T attack Hawaii with Japan in the next match ... so I do like the randomness of this AI ... it's very unpredictable and challenging and that randomness element to it is needed and glad you include it for sure. I know 1 time the AI could have taken my Russian capitol when it was on the ropes and could have sent in a few Planes and a couple Tanks vs my 1 lone UK Plane holding the Russian capitol but the AI decided not to and that was weird but it did eventually win. I know a couple times the AI left its Russian AI capitol open to be taken and I was surprised and my last game save above has 1 example of that but overall thats all I remember that is worth mentioning but pretty much theres nothing I can think of to make the AI even better since it just kicked my ass many times over but wow it was fun and look forward to the next round of testing!!




Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

Cernel
The big jump for v1 Allies AI is most likely the fact that redrum recently supported v1 placement rules in original factories (an enhancement I suggested him a while back).

Previously, the Russians AI placed 3 infantries in Karelia and 4 infantries in Russia, on round 1; and retreated all forces from Karelia, allowing you to take Karelia and round 1, which is terrible play.

Now, instead, the Russians AI places 7 infantries right in Karelia and stacks it.

So, no wonder that Allies now are much more challenging. Up until the v1 placement was not supported, letting Russinas be played by AI was pointless, since placing all in Karelia is so much important that not doing it just wastes the game.

Also, as I said, v1 is highly unbalanced in favour of Allies (usual bid is 18~21 for Axis, or skipping first combat move with Russians). Allies are normally harder to play, so all previous (very bad) AI were actually better with Axis; but the more the AI will get better, the more the map will be unbalanced in favour of Allies, even when played with AI.

I think that if redrum adds a good support for multi player attack/defence/coordination, then the map is going to definitively become hard to be played as Axis against Allies AI (with no bid). Anyway, when playing as Axis it is fair to give yourself some bid; and when playing as Allies it is fair to skip the first Combat Move (only) with Russians (not the whole first turn, only the Combat Move of the first turn).

Basically, if two very good players play v1 standard (no bid and no skipping of phases), the Allies player is supposed to win about 95% games with dice and 100% games with Low Luck.

I had a talk with "unbeatenclassic" (the best v1 player of TA) (you can join lobby and have a match with him), and he actually said that he thinks that the balanced bid should be 24 for Axis, or even more, if both players are very good.
History plays dice
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

captaincrunch
This post was updated on .
Thanks Cernel, that makes perfect sense. I more just posted to see if theres any way to improve an already good AI. I posted my game saves of my wins and mentioned anything significant and the most significant was 2 situations where the Russia AI stacked more offensively to take Germany that I was playing and left its capitol bare and too bad I didn't save the one situation where the Russia AI left its capitol empty and I could take its capitol for free with a Tank ... I am sure those are the situations Redrum could use to analyze improving the logic. Its an evolving thing. I actually am looking for the official launch of this game so I have some kind of closure hehe ... all fun.

Games Without Frontiers - Peter Gabriel;

https://www.youtube.com/watch?v=3xZmlUV8muY

Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

Black Elk
In reply to this post by redrum
Still enjoying that hard Ai Axis. They play well pushing into the endgame. Most of the goofs I see these days are just turn order exploits, leaving defensive fighters exposed etc. Otherwise hardAi has been stacking pretty well. The Japanese dedication to Alaska really has a way of drawing out the Allied game in 1942.2 such that it lasts several more rounds than it might otherwise. But this works to the strong advantage of Russia, if you are willing to just grind it out and push fighters towards the center for a dozen rounds or so. I like Japan's attack route against North America though. It's one of the things that keeps me coming back, just for the amusement.

I think I will almost be disappointed if the hardAi ever figures out how to just drive against Moscow full steam haha! The novelty of the Japanese North American stack invasion still is not lost on me, though I've seen it play out now several times.

Here's a game where Axis did pretty well, right up until they didn't ;)
Mainly due to turn order exploits, first in Karelia, and then more critically in Belo, which eliminated Axis airpower at a go.

HardAI_Allies_vs_Elk_Russia_Belo_fail_round_10.tsvg

HardAI_Allies_vs_Elk_Russia_Belo_fail_round_15_KGF.tsvg


For a randomized element, similar to the way purchasing is randomized, it might be fun if each "team" just chose an enemy capital at random, to throw the weight of their forces at, just constantly driving. In the case of 1942.2 for example Hard AI Axis might decide on Moscow, Washington or London as the main objective at random, and then keep to it for the duration. Hard AI Allies might pick Berlin or Tokyo and again throw the majority of their attack power against that single Nation. I think that might be novel for an AI attribute. Like sometimes Japan just guns all out for Moscow, or sometimes they try to take as much from UK as possible, other times they wade heavy into North America, but keeping up this behavior of several rounds.
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

captaincrunch
Where'd Redrum go? On vacation or maybe givin his brain a rest or maybe astigmatism from starin at the screen so much hehe ... anyways I'm just watchin for updates
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

Lasse i Gatan
In reply to this post by redrum
It is time for me to give some feedback again. I have continued testing new updates to the hard AI, but mostly on a map I have created myself. The map is first created with Random Map Generator, then tweaked to be pure v3-rules and units, with national objectives added to each nation.
I am uncertain how to upload savegames to help with the feedback. I think I have to, along with the savegames, upload the Random_Campaign folder too ( with the map in the subfolder games). Because my map uses icons in Random_Campiagn. Is this correct?

Feedback:
Purchase: Very good mixture of unit purchases! Perhaps the AI likes bombers and battleships a bit too much, but it is not a big deal.
The AI has a tendency to buy more transports, rather than using existing ones (as "ferries"). Something to put on the improvement list.

Combat: Very good! I have nothing to add to the improvement list. Only my own request, that I have mentioned before: Consider make the AI able to launch an attack even at unfavorable odds (but not at unfavorable TUV). Playing without low luck, I now that sometimes a reckless attack pays off!
A reckless attack also add more action to a game. It is a bit boring when two AI nations build up a large stack respectively, each afraid to attack the other.

Non-combat: Good! I do not like that capitols are abandoned just because the odds are against defending the territory. That is especially true, if loss of capitol also means loss of cash (to the captor). If so, the capitol should be defended at all cost.
Landing planes on carriers is already on the improvement list. Really hope Redrum can make it happen!

Place new units: Very good! I see no issues on my map.
Reply | Threaded
Open this post in threaded view
|

Re: AI Development Discussion

captaincrunch
@Lasse i Gatan, to upload a gamesave when you post, there is 9 options above your post when you are making it and its the last box called "More" and click on it and theres a "Upload a file" option and then find your Triple A gamesaves in a folder and click on it and then also click on "Upload file" and the gamesave will go into your post.
1 ... 34353637383940 ... 63