Current AI shortcomings I would like to see corrected/improved by a hypothetical Machine Learning type AI model
New
Okay folks, rather than me (an AI know nothing) trying to explain how such an AI might work. I’m going to present the shortcomings of the current models so people understand what it is I actually would like to see. Sorry if i presented LLM as some kind of panacea, I am just as curious as I am ill informed.
Disclaimer: Many of the issues I bring up here are going to pertain to Hearts of Iron IV. This is because IMO they work as better examples that more clearly demonstrate the problems of in game AI. And because as of right now I have much more playtime with HoI4 than with Vicky3 so am better able to comment.
Issue no. 1
The HoI Front AI logic has basically zero ‘operational’ to ‘theater’ level strategic understanding. For example, the German and Soviet AI’s will only ever conduct entire front length general offensives. I’ve never seen them concentrate forces to achieve the local superiority that would enable a successful advance. If their forces are stronger they grind forward, if equal they stalemate, and if inferior they lose ground.
If you concentrate a bunch of troops in one area, they do appear to try and match your force in that spot. But this is purely reactionary. In my own Eastern front campaigns I am able to strip my frontlines bare with no reserves to concentrate for an offensive. If I was playing against another player you can bet your bacon they would throw some armies at these naked lines just to force me to scatter my own troops to counter them. The AI is not capable of calling this bluff. And even when they sort of do, it’s not to the extent to actually be dangerous to my lines. Lose a few provinces and consolidate. No big deal.
Issue no. 2
Similar to issue no. 1, the AI has no understanding of encirclements. They pull them off themselves by accident sometimes but never coming close to what players can pull off. On the defensive they are even worse, there are times where I grind my way through their territory with a downright huge pocket forming containing dozens to 50-60+ enemy divisions, and a vulnerable skinny tether of territory connecting them to the rest of their lines.
On a local level the AI will shuffle troops into the relevant provinces to hold the line, but without even a fraction of the urgency they ought to have. A competent player/AI would immediately realize the defensive positions in the rapidly forming pocket are utterly untenable and would throw everything in and near the pocket to holding the lifeline and/or breaking out of said pocket should it be closed. To the AI, that inside of that pocket is just more frontline.
Issue no. 3
The AI doesn’t really grasp the importance of the supply infrastructure. I’m guessing supply depots and railroads have more “weight” assigned to them than a typical province. But very often I find myself cutting a critical enemy railroad that maroons a huge portion of their front, without them throwing everything but the kitchen sink to stop me or take it back like I would.
So while I appreciate the better supply system, it ultimately tilts the playing field further against the AI vs a player.
Issue no. 4
The AI is unable to properly concentrate or use armored/motorized units. They are spread across the front with no attempt made to concentrate them, and really get their money’s worth out of said expensive forces. This is ultimately an extension of the lack of strategic vision I highlighted above.
Issue no. 5
The building AI logic in Vicky 3 is pretty blind to what would actually be useful to build. There are countless threads discussing this. But a personal example is playing as Sweden. The market price of sulfur was through the roof, and despite this Norway was happy to sit atop its sulfur deposits not bothering to develop at all beyond a single level in a single state.
Issue no. 6
The diplomatic AI in Vicky3 will often charge into unwinable plays/wars due to their relatively simple weighted values. Prussia dashing itself against the rocks is a recent one I read about in another thread. Because of their own AI template, and the reckless value assigned to their unification play they will happily pick a fight they can’t win instead of bidding their time to gain strength, and allies. Or taking advantage of their adversaries being involved in other wars.
- - -
So it may be that machine learning algorithms are completely unsuited to solving these sorts of problems, or such solutions are a decade away. But my dream is to feel like I am playing a multiplayer match in single player. I am very busy with my 2 month old baby, have a wife to spend time with, and work full time+ hours so actual multiplayer isn’t really practical for me.
Disclaimer: Many of the issues I bring up here are going to pertain to Hearts of Iron IV. This is because IMO they work as better examples that more clearly demonstrate the problems of in game AI. And because as of right now I have much more playtime with HoI4 than with Vicky3 so am better able to comment.
Issue no. 1
The HoI Front AI logic has basically zero ‘operational’ to ‘theater’ level strategic understanding. For example, the German and Soviet AI’s will only ever conduct entire front length general offensives. I’ve never seen them concentrate forces to achieve the local superiority that would enable a successful advance. If their forces are stronger they grind forward, if equal they stalemate, and if inferior they lose ground.
If you concentrate a bunch of troops in one area, they do appear to try and match your force in that spot. But this is purely reactionary. In my own Eastern front campaigns I am able to strip my frontlines bare with no reserves to concentrate for an offensive. If I was playing against another player you can bet your bacon they would throw some armies at these naked lines just to force me to scatter my own troops to counter them. The AI is not capable of calling this bluff. And even when they sort of do, it’s not to the extent to actually be dangerous to my lines. Lose a few provinces and consolidate. No big deal.
Issue no. 2
Similar to issue no. 1, the AI has no understanding of encirclements. They pull them off themselves by accident sometimes but never coming close to what players can pull off. On the defensive they are even worse, there are times where I grind my way through their territory with a downright huge pocket forming containing dozens to 50-60+ enemy divisions, and a vulnerable skinny tether of territory connecting them to the rest of their lines.
On a local level the AI will shuffle troops into the relevant provinces to hold the line, but without even a fraction of the urgency they ought to have. A competent player/AI would immediately realize the defensive positions in the rapidly forming pocket are utterly untenable and would throw everything in and near the pocket to holding the lifeline and/or breaking out of said pocket should it be closed. To the AI, that inside of that pocket is just more frontline.
Issue no. 3
The AI doesn’t really grasp the importance of the supply infrastructure. I’m guessing supply depots and railroads have more “weight” assigned to them than a typical province. But very often I find myself cutting a critical enemy railroad that maroons a huge portion of their front, without them throwing everything but the kitchen sink to stop me or take it back like I would.
So while I appreciate the better supply system, it ultimately tilts the playing field further against the AI vs a player.
Issue no. 4
The AI is unable to properly concentrate or use armored/motorized units. They are spread across the front with no attempt made to concentrate them, and really get their money’s worth out of said expensive forces. This is ultimately an extension of the lack of strategic vision I highlighted above.
Issue no. 5
The building AI logic in Vicky 3 is pretty blind to what would actually be useful to build. There are countless threads discussing this. But a personal example is playing as Sweden. The market price of sulfur was through the roof, and despite this Norway was happy to sit atop its sulfur deposits not bothering to develop at all beyond a single level in a single state.
Issue no. 6
The diplomatic AI in Vicky3 will often charge into unwinable plays/wars due to their relatively simple weighted values. Prussia dashing itself against the rocks is a recent one I read about in another thread. Because of their own AI template, and the reckless value assigned to their unification play they will happily pick a fight they can’t win instead of bidding their time to gain strength, and allies. Or taking advantage of their adversaries being involved in other wars.
- - -
So it may be that machine learning algorithms are completely unsuited to solving these sorts of problems, or such solutions are a decade away. But my dream is to feel like I am playing a multiplayer match in single player. I am very busy with my 2 month old baby, have a wife to spend time with, and work full time+ hours so actual multiplayer isn’t really practical for me.
Last edited:
- 2