The great official AI thread

[BjornB]

Please don't post feedback collected in 1.2.1 in this thread! We'd love to hear 1.3 feedback though.

And make sure you post detailed explanations on what aren't performing well.

To quote SteelVolt:
"Post your constructive AI criticism and feedback in this thread. Discuss what frustrates you about it in a civil manner, please

So, if you have something you want to lift in this thread, I would like you to include the following information:

• What the AI is doing wrong
• What you think it should be doing instead.
• Any potentially relevant context, as many things when are crazy in one situation is a perfectly reasonable thing to do in another. For example, if Japan does not invade USA it may simply be because they are getting their ass handed to them by Soviet."

Non-constructive or outdated feedback will be removed to make this thread useful to us.

Thank you!

 

[SteelVolt]

Hello everyone!
This is your friendly neighborhood AI developer (as many of you no doubt already knows), and we need to talk.

Even though the discussions following the recent AI dev diaries, I think we all know that it had not always been so positive.

Believe it or not, but it has not gone unnoticed that several of you have strong feelings about the AI in Hearts of Iron 4, and I am certainly the first to agree that there are both bugs and room for improvements. However, derailing threads (or indeed the general discussion) with comments along the lines of "AI is shit" is a waste of everyones time. More yours than mine, as it takes you more time to post it than it takes me time to scroll past it. And it certainly does not encourage me to talk about what I am working on.

Before you take this as an attack, I know almost all of you are really positive about the game, and even those who are less positive are still good and reasonable people even when some of you let you passion get the better of you. I get that a lot of the feelings stem from a desire for the game to be as fantastic as it should be, and I share that desire.

Bottom line is, I am starting this thread to try to drag the discussion out of the sewers it sometimes ends up in, and I would ask for help from all of you in doing so. I do welcome constructive criticism and feedback, and henceforth this should be the place to do that, with the notable exceptions for suggestions and bug reports, which both have their respective bug forums. One of the reasons I do not get involved in discussions about the AI is the toxic tone, which I will simply not encourage by giving it attention.

Not everyone is going to agree with all design decisions, and that is fine. Just remember that if I do not agree with you it does not mean I am not listening.

So, to kick things off I thought I would give you all some points about the status of the AI work on HoI4. And even more insight can be gained by reading the AI dev diaries preceding the Together for Victory launch.

First of all I would like to point out that HoI is one of the more complex strategy games out there, and we are a relatively small team developing it. I am, and will continue to be for a foreseeable future, full time on AI. We are also working on getting more developers involved and capable to work with the AI.

Given the complexity of the game, writing general purpose AI is not like making any of the features of the game, which have clearly defined limits. The AI is more like a work of art in this sense.

The hardest part of building the AI is not making the reactions to situations, but detecting the situations to begin with. Typically due to that it often comes down to pattern recognition with more variables than it usually appears to a person. Humans are extremely good at pattern recognition.

There are a ton of issues on my list relating to almost every part of the AI, and aside from those I ave a ton of ideas for future potentially interesting and advanced improvements.

I will not be giving tons of details about the inner workings of the AI, or exactly how I am solving problems, mostly because that would take too much time for me which would otherwise be spent working on the AI.

Oh, and lastly: No, self learning AI is not a viable approach. Googles success with AI playing Go does not mean the problem has been solved as such. First of all it was trained on a supercomputer, and second; in terms of complexity, primarily number of parameters and potential decisions in any given situation and the fact that HoI is not turn based in the way Go is; makes Go look like tic tac toe in comparison.

Tl;dr: Post your constructive AI criticism and feedback in this thread. Discuss what frustrates you about it in a civil manner, please

So, if you have something you want to lift in this thread, I would like you to include the following information:

• What the AI is doing wrong
• What you think it should be doing instead.
• Any potentially relevant context, as many things when are crazy in one situation is a perfectly reasonable thing to do in another. For example, if Japan does not invade USA it may simply be because they are getting their ass handed to them by Soviet.

 

[podcat]

Torch Update: I´ve just let a hands-off game as bulgaria run till late ´40. First of all, the improvements in AI are significant, because germany is very capable of taking poland, the low countries, france and denmark in time..

Nice to hear

 

[podcat]

I've run 2 hands off games. In both, UK capitulates Italy by 1941. This forces Germany into a brutal mountain war with the UK. Now I know this is only 2 games and this is a pretty cool thing to happen on a rare occasion but 2/2 is probably too often.

Yes, its something we have noted is happening a bit too often are are working on improving. Its also a bit of an issues on the allied side since doing this too early often make them bleed more than they should

 

[BjornB]

@BjornB You deleted my post (a post that have taken me some time and effort to write) but I'm not sure production, template and variant issues are fixed in the new patch.

Can't see how it wasn't constructive.

Okay, undeleted since you obviously put effort into it. But maybe you could copy the text, play some 1.3 and make potential adjustments (if needed) and re-post it as 1.3 feedback?

 

[BjornB]

I can do that. Let me copy it.
Thanks.

Edit: Done. Thanks again.

Great, thanks

 

[podcat]

....podcat? what?

yeah this is SteelVolts thread. Hes gone home for the night now tho, but he should be back tomorrow morning with replies

 

[SteelVolt]

USA joined the Allies in 1942, rather late but hey Historical Mode. USA had 500 Division to trough at Germany, but nah. Let's sit on half on my own soil (this is ok) but let the other half stay in transport on the atlantic, and don't move them. But let them eat 40% Attrition (Most certainly a bug)

Yeah, we are aware there are some issues with both when and how USA gets involved, so it is something that will be looked in to.

France shouldn't contest Rhineland

Indeed, in historical mode this should not happen. Sounds like a bug that will need looking in to.

1). Invasion preparation - Germany and Japan especially. When these two countries (under AI control) declare war the vast majority of the time they have no troops on the border with the enemy or they seem to have a weak force at the front that they then try to flood with reinforcements and it turns into a stalemate with low supplies everywhere. I have no idea if it's possible to program the AI to have "forethought" about how it wants to invade but that would be one request.

Wow, other than when they do not have a border of their own to an enemy, I have not seen this happen. There are typically early ques for when a war may be happening, such as focuses that would lead to it or wargoals being generated, and the AI does look at these. It may be issues in the front balancing unrelated to wars as such, and if that is the case it is something I have kept working on even after submitting the 1.3 build, so expect some improvements already in next update. If you keep seeing this, please report it in the bug forum.

2). Usage of military power - This one bugs me the most when I am doing all the major lifting in a fight and I see my ally just sitting around with a few forces at the front and the majority of it's military strategically redeploying around the world. It's very obvious why it's doing this when you tag over to the AI and see the front lines that it has created that it could never cover or on fronts that don't even have enemies. I really really hate seeing German and Italian division fighting in China when Poland and France have not even collapsed yet.

The moving around of units is a continuous uphill battle, but I keep chipping away at it. When it comes to creating reduntant fronts during war, I have noticed the same thing myself. Sorry that is the situation in the current version available, but I am already working on it

3). AI Military Action - This one is a reason why I have stopped playing Ironman. Example: Playing as Germany invading the USSR I push and capture pretty much all major victory points and USSR is at 98% surrender. I get Japan into the war so they can capture Vladivostok and end the war. However I watch in horror as the Soviets have 1 or 2 divisions guarding Vladivostok and Japan as 8 divisions sitting by but does not attack. My men are dying for empty ground that has no value waiting for Japan to move just a little bit to the east to end the war but they don't. I have to tag over and make them attack to end the war. Two way I see this can be fixed. One is the AI being more aggressive especially when the surrender percent is so high or make it so the USSR will collapes after Germany has conquered everything west of the Urals.... I was waaaaaayyyy beyond the Urals moving into Mongolia and Siberia and USSR still would not surrender until I made Japan take Vladivostok.

I feel your pain, friend. Balancing AI aggressiveness is something I keep coming back to, and will keep working on. Thank you for a good example of where this fails. Please feel free to report this in the bug forums with a save of the situation. If not, I am still on it

 

[SteelVolt]

- What is the Italian a.i. doing?
It just dow'ed Yugoslawia.

- Why is that wrong?
Uhm, Italy is knee deep in fighting France and trying to contain an invasion near her capitol.

- What should Italy be doing?
Priorities! Italy should focus on surviving before opening another frontier.

Indeed, that is much too brave. If you have a save from recently before they dow Yugoslavia, please report it in the bug forum. Either way I will keep my eye on it.

 

[SteelVolt]

Podcat, thanks for making this thread.

- The AI should stop shuffling troop everywhere. This an obsession for the AI. It should decide on an acceptable force distribution and stick with it for the most part.

- When I order the AI to attack please for the love of god move forward. Not sideways. Not backwards. Not even really diagonal. It should move forwards to the offensive line I have instructed it to take. Don't shift units around.

I have kept working on point one since we submitted 1.3, so expect some improvements to that when we push the next patch (I do not know exactly when at this moment).

The second point is something I am painfully aware of. I will absolutely be looking in to how this can be improved in the future.

 

[SteelVolt]

Lol, oops. I saw the avatar and just assumed it was podcat!

The love for error dog is universal.

 

[SteelVolt]

@SteelVolt good to have one thread going.

seems btw 1st game as Sov, after Germany has occupied France for a while i thought i'd have a look at its Resistance management..

and its still 0%

maybe the AI should get 'dispensation' from it to not get so much damaged stuff it cant seem to control
among infra damage , usually its factories (Germany in this case where it hurts alot) are 16-33% damaged from sabotage alone , maybe it's an idea to make resistance have no effect on any AI nation until they can learn to manage it?

Darn it. Thanks for heads up. While it was not something I had been focusing on, I was hoping it had gotten better through other fixes.
Anyway, it will for sure be looked at in due time.

 

[SteelVolt]

Tha AI is better but still have problem for priority front. I go for little entente the war began for the Czech and Italiy join the fight, but the Ai think it is better to move +/- 20 divisions of the French frontier to Yougoslave one will I'm ready to attack, it react fast so I don't go too far but i take Milan with minor fight. Same for Germany, it put troop before DoW Czechoslovakia but keep only 4 divisions for the maginot ligne.

Thanks for the examples. I am aware, and the work on this has been and is moving on since we submitted 1.3. Some improvements should be in the next patch, with hopefully more to come.

 

[SteelVolt]

I just played as Ethiopia in the 1936 scenario and managed to free all of the Ethiopia home land from Italian forces in March 1937. Overall Italy didn´t act agressive enough with its much superior forces but still I was almost lost despite having taken one of the two important ports. Then Italy forgot to guard its only port left in its own advance and soon completely crumbled due to missing supply. Not understanding port supply seems to be the biggest issue of the AI right now.
Apart from that the AI seems to be somewhat improved, both the player AI as the opponent AI. At least I had the impression when I played Germany against Poland in the 1939 scenario. My German troups under AI comannd took Danzig quickly and the Polish AI tried to mount counter offenses to East Prussia as long as they could.

I found a bug in how the AI set up groups just the other day, which had a massive impact especially on Italys Ethiopia setup. Improvements should be coming in next patch or update (a fix is for sure in the pipeline.)

 

[SteelVolt]

• What the AI is doing wrong

when assigning factories to production, the AI treats planned (grey) and inoperable (red) factories as if they're perfectly functional. Frequently leads to situation where the equipment the AI needs doesn't get produced for months or even years, because it was assigned nothing but such unusable factories, and the AI never bothers to correct it.

• What you think it should be doing instead.

when distributing factories for production, only take into account what's actually available and working ('green' factories/docks) That you think you'll have some more few months down the road is nice, but maybe worry about assigning those once you have them up and running.

Thank you. The damaged ones is an oversight that needs to be addressed, but the assigned grey ones should not happen other than temporarily when they get occupied. The AI should correct it daily. Really seems like a bug to me. Could you please report this in the bugforum?

 

[SteelVolt]

Long time lurker, first time poster.

Is this the place to complain that in an observed game, Germany only had FIVE tank divisions by May '38? I also have an ironman save which I suspect suffers from the same problem, as I can't find ANY German tanks by '40 something, and I retook the lowlands and the Maginot line with divisions composed of 50/50 artillery and inf.

I've been lurking in the forums since HOI came out, so I'm aware of the AI drama, the Blitz command things, etc, don't need to be brought up to speed on all that, but the German AI seems 100% NOT improved at all since Sunflower.

Just wanna know who/where to send these save games to....

Thanks, love all the games otherwise.

------------------------------------------------------------

And, to comport to the thread rule request:

What the AI is doing:
Germany isn't building tanks almost at all, the rest aren't building enough.

What I think it should be doing:
Build tanks.

Not trying to be even slightly flippant.

You have indeed come to the right place. When it comes to buildup, saves are rarey helpful. The German buildup is something that has been looked at quite a bit, though not all improvements were completed before we submitted 1.3. So expect some more improvements to this soon. If, however, you are not seeing any improvement at all since 1.2.1, You may want to make sure you are not running any mods that could be related to production/deployment that were balanced for 1.2.1.

 

[SteelVolt]

-The Pacific isn't as contested as the AI patchnotes said they would be. It is still a calm ocean.

It depends a lot on what else is going on. It might be that the concerned countries have had their capacity to mount a pacific war beaten out of them elsewhere? Or just that other theaters are more critical?

-Britain is fabricating a claim on me, and I am guessing it is because I generated 3% world tension by capitulating Manchukuo. Note, I am still part of the Allies as India, and used the new autonomy system to become independent.

Is it an actual fabrication or is it through a focus? The allert makes no difference between those two.

What the AI is doing wrong: Italian troops in Africa are in the deep Libyan desert, taking attrition, when there are no enemies opposite of them. Supplies are terrible and they take about 40% attrition. Some units are standing still, others are shuffling around along the "frontline".

What the AI should be doing: Avoid these high-attrition/low-supply zones unless there are enemies opposing them there.

Just noticed this while playing now, I hope this helps.

Holy hell, I have not seen that many Italian in north Africa in a long time. That is clearly an overly ambitious for a desert.

I'm not sure if it's a 'bug' per se, but strategically it doesn't seem to be in Germany's interest, particularly with so little of Poland taken. Have attached a save in case useful.

I would really need a save from before Germany declares war to see what tips them over the edge. Unfortunately we have focused our tests on historical, so these types of situations that would not happen in historical has flow under our radar. I appreciate the heads up.

 

[SteelVolt]

Also in my Historical game, Poland joined Comintern instead of Allies, which left the situation problematic for me being the UK, leader of the Allies. I could not invite Netherlands or Belgium until they had been invaded and once the world tension was above 80%, and then they had to call me and France into the war. Correct me if I'm wrong but 1.2 was not like this? Surely if you're in the same faction you'd be automatically pulled into the war? There seemed to be a problem with World Tension not accumulating as quickly as before too.

Poland joining comintern in historical? Without UK player blocking them from Allies or a Soviet player actively trying to get them in? Do you know if it is something than can be reproduced?
I have not seen this happen without player interaction, both in 1.2.1 and 1.3.

As for being pulled into war: before an automatic call to arms was sent, but could be declined. Now you have to send it manually. This is because certain levels of subjects are not allowed to decline call to arms, so the master decides when and if they should join. This meant the the automatic call to arms needed to go in other cases as well.

• 
  
  
  • What the AI is doing wrong

Existing trade routes that have trade relations changed to below zero will still send convoys but to no avail. The convoys still go out but bring nothing back.

• 
  
  
  • What you think it should be doing instead.

Instead of sending out convoys for 0 gain, the trade route should be cancelled.

Context:

Create a trade route with a country then lose trade relations with that country. Your trade route will still exist and the convoys will be sunk but it will bring back nothing. Also you can't manually cancel the trade route.

The AI should adapt their trade deals if they are not getting enough goods delivered. Do you have a case where it does not?

My pet peeve with the AI is an old one, Not responding fast to a naval invasion.

With the AI priorities in the latest DLC, its possible to make pure coastal defence armies that the AI will happily spread out to cover the coast. However what happens is the enemy will launch a Naval invasion into a port/province, only the units stationed there will fight. The neighbouring units will not respond at all and just stand there. So imagine a single 'garrison' division defending against 6+ enemy divisions, fight goes on for a few days and eventually they lose. Throughout this period, the friendly divisions from the same army, controlled by the same AI will just stand and wait at the neighbouring province and don't even charge in to reinforce. They will only respond when the port is taken but usually by then its too late as the enemy can easily pour fresh units into the ports (they actually reinforce invasions properly now).

My request is, the AI has a priority to respond to invasions by actively moving neighbouring units (from nearby provinces) into a province that is Losing the naval invasion. If the AI has excess units nearby, they should throw more units into that fight even if its comfortably winning. Aka, I want the AI to not be Hitler and let Rommel reinforce the beaches, its too passive currently.

I agree. We are aware of these issues and are going to look at some ways to improve this.

The naval invasion is bugged.

The invasion mechanism might be tweaked but it means now that AI are sending everything, ALL THE TIME. Playing as China, Japanese AI keep sending small stacks of landing parties that goes nowhere and dies, over and over again, even when it only has 5 divisions left.

It's just getting ridiculous to see them getting ashore with 1-3 divisions when they only have 6 while I have 200+ divisions.

Maybe have the AI 'turn off' the naval invasion mechanism when condition triggers?

Thank you. If you have a save of the situation you describe it could help us to efficiently work on the balance for this. If so, please report it as a bug in the bug forum

Italy is still the weak point of the Axis. Surrendered in 1940 (!) in my game with Egypt. It also seems, that Germany can't adopt to this situation. They helped Italy to win back their main land and this delayed the fully conquest of France for nearly one year, although Germany advanced normal with their focus. In the end, Germany declared war against the Soviets, despite France was back again, half of the divisions were in Africa (Italy...) and Nat. Spain was nearly defeated. Oh, and Italy declared war against Yugoslavia a little bit after this.

Another problem:

AI is camping in the desert and eating horrendous attrition, even, if the front is clear. Same with the hills in Iran, Afghanistan and the others.

The attrition thing is something we are aware of and will be working on some time in the future. Italy is a weak point we are aware of. They really should care less about Africa now, but as I have seen others mention this I think it is worth a second look.

 

[SteelVolt]

1.
Problem: (German) AI declaring war after focus finished.
Expected: (German) AI not declaring war after focus finished.
Details: 1940 jan, Ger finihed off poland in '39 nov, starts to push against France and Benelux, Italy struggles in africa (playad az SA), ger losses already over 500K, Ger starts war focus against SU (ahistorical run), and starrts war ASAP. The eastern front not even manned, SU takes 2-3 province at the start of the war, later Ger slowly pushes back, but lost lot of troops and stuff.
The AI should choose better, when to go war, not choose wargoal focus "randomly", and even if it choose a wargoal focus, the AI should get a warning, this focus will generate war with country X, so it can prepare for that war during the 70 day of the focus.

2.
Problem: Not defending against air superiority
Expected: Defending against air superiority
Details: Same game, after Ger declares against Yuoslavia, sent my entire airforce (38 planes), to run air superiority in the balkan area, zero enemy fighters or planes whent against it for the entire duration (6+ month), until Yugoslavia fall. Later I sent the same 38 planes to help out in italy, and once again, no enemy airforce.
The Ai should counter small airforces, and even more, the allaid AI should help with his airforce in similar situations for both side. UK had 2000+ planes, France 500+, Ger around 1500+ and Italy 800+; yet less then 1000 planes used together.

3.
Problem: Defended semi-friendly boarder
Expected: Abandoned semi-friendly boarder
Details: Same game, china fights to the death agaisnt Japan, yet it fully maintains boarder guard against France, Raj, and SU, all three has non aggression pact with china, all in war against Japan. The AI should not leave boarders empty, unless they are friendly, or at least not opposed. Of course, the other side also manned the boarder, so they are also lost dozens of divisions to the unlikely event of pact breaking and backstabbing.

1: Assuming the focus script is set up properly, the AI does get notified that war is coming. If you have situations where this does not seem to be the case we would love a bug report on it. That said, there does seem to be situations where the AI is much too eager to declare war once it has a wargoal. Typically I do see it prepare a front before declaring, so if you have counterexamples I would very much like to debug a save if you have one before the crazy declaration.

2: Air AI being somewhat sub par is known and actively worked on. Expect improvements soon

3: I tweaked this just the other day, so it was way too late to add to 1.3. We need to look a bit at how the improvements I did to it effects balance before we can put it in a patch, though. But it is known and absolutely being looked at

 

[SteelVolt]

Aye, sorry, I just kind of fired it up and had at it - next game I'll turn ironman off so I've got the autosaves and'll be able to load up both. Is it more useful if I stick to historical mode or is getting awareness of these things in non-historical mode a good thing? (I'm happy playing either, and mix it up, so very happy keeping an eye out for this kind of thing in either).

No worries. As far as I am concerned it is our job to fix the game and you should just get to enjoy it
While I would LOVE to spend a lot more time working on ahistorical, we do currently prio getting historical up to a high enough standard. That said, when things go wrong in ahistorical it is still a symptom of something that COULD go wrong in historical. So I would suggest this: play whatever you feel gives you the most enjoyment, just remember to metion which it is when you report issues