It would probably be most useful to focus on a single AI country and compile as much of its activity as you can for each day in a spreadsheet. I'd pick a country that can't attack me and is least likely to be attacked by another player and use as many spies as you can afford.

I am very interested in this i do very much hope you can continue it to a successful conclusion

I've observed a few things about AI behavior. Curious to hear what others have seen, and what your experiments with Mexico show.

The AI operates its military on a province level. It has no strategic objectives.

When you place units on the border of an AI province, the AI will attempt to match the numbers. There is some kind of local army strength to opponent army strength ratio for every province. Provinces that are under greater pressure draw units from provinces that are under less pressure. For example, if France places a large stack next to a German AI, and Poland moves his army East, away from Germany, then German troops will over time migrate West to counter France. You can exploit this behavior by forcing the AI to vacate the border you want to invade, and then move in quickly to take over large swatch of lightly defended territory.

The flip side of this province level, power balance logic, is that the AI will try to take advantage of empty provinces. This is why fighting the AI along a large border can be so frustrating. As soon as you stack your units to push forward, the AI will stream across the border, threatening empty provinces with anything it has on hand. Then you either have to divert your artillery and air power to knock them out one by one, or suffer unnecessary losses in small, pointless, but bloody close combat all over the map.

The worst case scenario is when you completely surround an AI country, which wasn't a threat all game, but then you get large enough so that every AI declares war on you. That small AI country in your middle will now send units in every direction. This kind of mess takes a lot of effort to clean up. Tip: Don't let it happen. Kill everything while your front line with that country is still manageable. EVERY country is an enemy, in the long run.

The AI has a concept of unit type.

When the map is swarming with airplanes, the AI will start building more AA guns, and moving them close to your air base. This is assuming the AI has access to the tech and the factories. On maps with less air power, the AI seems to build less AA. Similarly for tanks and AT guns. I'm less sure of this observation. It may be accidental. More research needed.

The AI tries to balance its resources through the market.

It doesn't change its tactics depending on resource availability. Neither do most players, so this is not a knock on the AI's programmers. When the AI is lacking in food, for example, it will bankrupt itself trying to buy more, selling everything else in the process. Because of this, the AI is typically broke. Take cities (especially capitals) while they are still player controlled, else your conquests will not be very profitable.

One interesting outcome of this trading behavior is that food-rich AI countries will tend to do the best. They will build effective units (infantry, militia, AT guns) in greater numbers. An AI rich in rares will end up with more flying bombs, or nothing, because it's trying to buy food by selling its rares.

The AI researches tech in breath-first order.

Meaning, it will try to get every tech to level 1, then 2, etc. It doesn't go deep into any tech tree. Again, noob players do the same thing. This is not a knock on the AI programmers. But this leads to bad strategic outcomes. Instead of fielding level 2 AT guns to match your medium tanks, it will be researching flying bombs, because that's next on its research menu.

AI control over air power, naval power, are complete unknowns to me. There is no obvious logic to it. Sometimes, they will attack. Other times, they will patrol, or go somewhere else. Often you can simply go around AI navy stacks, take all their land, and the navy will disappear without a fight. The only issue is a battleship or cruiser close to the shore, making it harder to take that last city. Often you can just wait for it to sail away, and then take the city, so you don't have to spend a single HP.

My observations have been similar to yours, though I've also never paid much attention to them because they're so predictable and easy to defeat.

One thing I've noticed is that they are almost always moving units dot-to-dot, rarely sending units on longer journeys. This holds for their reinforcement behavior, where they'll draw units from the nearest couple provinces and fill them back in later. I've seen some instances of them sending units multiple provinces when there's an empty province to defend, and some rare instances of them sending units on long treks to the front line, but otherwise it seems like they're usually moving one province at a time (or one dot at a time for naval units).

I wonder what kind of data Bytro gathers and whether some sort of machine learning algorithm could watch players' decisions and learn from that to create a stronger AI.

I've thought about this as well. How would they select the training data? Which players and which games would they mine? They would have to pick players who doesn't use gold (because gold-based play styles don't work without gold). They would have to pick players who win regularly, and against top opposition. They could optimize for every doctrine. The end result would be much better at CoW than most human players. Only the best could compete against the AI. The noobs and golders would give up and walk away. It would kill the game

All good questions. They could segment the player population by win percentage or k/d and set up multiple AIs that game creators could choose from, but the default maps would probably have to be pretty basic level. Maybe up the AI competitiveness with map size, with the 100p map getting a harder AI than the 50s, which has a harder AI than the 22, etc.

Back to the discussion of AIs as they currently are, my observation has been that the AI makes multiple decisions at the same time--for example, you'll often see multiple AI units start moving simultaneously. I assume, but don't know for sure, that construction or production starts also occur at this time (maybe Python can confirm if they can get enough observation time). I wonder if the developers could increase the frequency of these decision times to make AIs a little more active and therefore more competitive.

Another unrelated idea could be to give AIs an economic advantage to make up for their strategic inferiority, like cutting all construction and production costs for AI by 25% or something.

Another random idea. Make the AI play better against better opponents. Give the noobs a chance, on every map, regardless of map size.

The poorly programmed IgnorAnt countries have no concept of unit types.
For example, they will build Naval Bombers in Moscow, with NO airstrips available to
get anywhere close to water.

This is true. I also have the sense that if you put more units near their border they are more likely to declare on you. Any thoughts? More importantly, what is your feeling on how they can detect you? Do you have to be in visible range for them to count your units? Fog range? Any range?

Yes! They don't do shit and then all of a sudden they all start marching to this place or that. There seems to be some timer where some calculation is made and everything moves. Individual stacks can move at any time though depending on immediate reaction, e.g. artillery.

Regarding planes that someone commented on, my experience is that they usually pick one of your cities to attack and keep attacking it, even if it has no units. This is easy to kill without losses by patrolling over the city.

Regarding AI. It would be pretty easy to use machine learning to dominate most human players. Obviously, bytro doesn't want to do that. They are not using machine learning for AI. It is just some simple hacky code.

the IgnorAnts have perfect knowledge of your troop disposistions,
they will attack an empty province far out of the observable range.

This is very useful
when at war with an AI
they only issue move orders every 12 hours, unless they are being bombarded

Did anything ever come of this?