Huh. Did you edit your post? Because I got an email reminder with quite different text.

But tracing the AI development to W2 is an excellent explanation why it's so simple.

KGB wrote:


I've asked for clarification, but my first guess would be, Light applies to 1-turn units. Flyers fly, grunts are everything else, or based on their combat stats, which flyers would then get an exception to. So a flyer with good combat stats for its tier, like a griffon, could also be classified as grunt.

I've gotten some answers on money earlier:



I'm not quite sure about that last part, to me, it would make more sense that freshly pillaged cities without any production should have a lower threshold for buying production, not a higher one. But either way, those are the figures Roger gave me.

I have asked him about the 1.02 code too.

In light of Roger's code annotations, I've added an 'AI value' column to my spreadsheet, both for my modded units, and the originals.

docs.google.com/spreadsheets/d/1pQCk7eml...825194#gid=563825194

Keep in mind, I've made a few assumptions which could still prove incorrect:

• Siege: It's interesting that this is the only power which gets a special flag. I've assumed Siege gets counted twice, both here and as a regular power.
• Power: I've assumed having any kind of special ability sets this flag, regardless of value. But it's possible this one only applies to a subset of abilities.
• Scout: Roger's note was a bit cryptic. I've assumed, either Hill or Forest bonus is enough to set this flag, Flying also applies, and there is no further bonus for having both Hill/Forest. It could be possible that other terrain bonuses also count, like Marsh or Desert, even if they're not active on the current map.
• Fast and Very Fast: Here the obvious conclusion is that these stack.

If these values are added like I've described, then Dragons get a score of 19. 1-turn units range between 7-10. In testing, Dwarf Crossbows were much rarer than Dwarf Infantry, despite having the same score. That might imply not all abilities set the 'power' flag, and Missiles isn't one of them.

But on the whole, it looks reasonable, and I've adjusted some unit stats or costs based on those insights.

Hey Tim,

I think it might be more prudent to continue talking here, in case it'll be of interest to other people.

Between correspondence with SSG's Roger Keating, and feedback from KGB, we have pretty much figured out how the AI evaluates its units, and what subcategories it's dividing them into when choosing one to produce.



This is still the core. What is important to note, if two units get the same result due to the random dice rolls, the lower slot in the army list gets priority. Between two units of equal value, that means the lower one has a 10/16 chance to be picked, the higher one 6/16. If the higher one is worth a point more, that flips. At two points, it's 3/16 vs 13/16. At three points, 1/16 vs 15/16.
At greater than 3 points difference, the only chance the lesser unit gets built is if the other one costs more gold than the AI has in its treasury.

SIEGE gives two bonus points, because it also counts as a regular power. Cities can be allocated to produce SIEGE troops, so even if their value is lower than others, these will be produced. That's why you see the AI use Catapults and Siege Towers, despite their low inherent stats.

As best we can tell POWER applies to any units with a special ability, which is almost all of them. There is no special production on any subtype of powers, or based on how strong the ability is.

SCOUT applies to units with have both active terrain move bonuses for the current map. For random, that's always Hills/Woods, but other combinations are possible in custom scenarios. SCOUT troops get priority for searching armies in hidden map games, though not as much as flying.

FAST means movement of 16 or higher. It stacks with VERY FAST. Some cities may be ordered to produce only FAST troops.

VERY FAST means movement of 24 or better. Like FAST, cities can be assigned to produce only VERY FAST troops.

Apart from these, the main classifications are:

LIGHT FLYER - A flying unit that takes only 1 or 2 turns to produce.
FLYER - A flying unit with at least 4 strength.
LIGHT GRUNT - Any unit that only takes 1 or 2 turns to produce
GRUNT - Any unit with 4 or higher strength.

A unit can have more than one of these flags.

What I've done is revise a few units between the 22/24 movement brackets, as well as 15/16 in a couple of cases.

I'm now once again pretty happy and confident in my current unit stats, but it will take some playtesting to notice any odd behaviour.

As far as adjusting random sides is concerned, I think I'll start over. Now that I know how the AI behaves, it's clear that it takes some careful curating to have an army list that the AI can make full use of, at least its regular troops. Thankfully, outliers that wouldn't normally get produced can always be put in mercenary slots.

The next big task on my to-do list is study vanilla hero leveling and spell-casting. Then, I'll adjust my heroes and spells based on those findings, so the AI will hopefully pick spells and abilities that are useful.