I'm not an expert -

To look for yourself, use the launch property -gpldebug, and push the (_) key during gameplay (and read the documentation). You'll want to scroll past all the things that start with a # sign, since those are mostly the integer values for various tasks. Hero AI is contained in things such as Wizard_tree, and Adept_Tree.

iFrom what I understand, Majesty uses decision tree based AI to simulate the hero freedom. they have a full list of all their actions with values from around 15 to 85. and as they pass or fail checks, they'll choose to do one thing or continue down the list of checks, reseting to the top of the list if they complete an activity. The very last check is the 100 value "wandering" which is only likely to be seen very early in games when heroes have nothing to do. It seems to act as a failsafe to reset their decision tree to the top in the event that all monsters are dead / player hasn't built anything the hero is interested in.

Different hero types weight different activities differently, and a lot of activities are further affected by modifiers (the most obvious being reward flags, which heroes don't go often as their high place in the list suggest due to being negatively affected if the gold on offer / distance is too low for the hero).

To preserve the game, all heroes also have a "subtask" which usually but not always involves scanning their immediate area - hence they are able to cancel most actions in respnose to treasure chests / combat / etc.

Ai for monsters and henchmen are more simple (although there are still 5~monster AIs) and focuses mostly on their pathfinding AI, ways to simulate "wandering" and getting them to start combat effiecently.

Thank you very much. I'll take a look at what you have said.

This can be considered a pretrained neural network?

Them switches between the primary and immediately subtasks around them, and they form automatic parties when adventuring.

I have take a short look at the debug screen, but not a deep look yet.

But thank you very much for helping me. I can use this as a form of "source" for my work.

I'm not sure majesty is advanced enough to be a pretrained neural network - It's not really attempting an artificial life simulation like Steve Grand's 1996 Creature game does.

I think there are two types of automatic parties

follow and support - no learning involved, rangers/healers/wizards simply have a strict pathfinding AI if they roll support_hero in their decision tree.

The 4 of same guild kind - (easily visible with paladin/warrior of discord) - I don't know the entire cause of this, but I believe it's the result of combat_wandering (shown ingame as hunting) when the hero pulls up the same target_eval result (probbably because similiar stats). These hero types seem to show the gameplay response of grouping together even if you get the the warriors of discord from seperate guild buildings.

Majesty does do a couple things to fake "learned" responses - the fight or flee calculation's use of hero damage output / health makes heroes flee less as they level up (and wizard seem to factor in range better), rangers with fire balm will know that it lets them fight skeletons who they would normally flee from, and in a few cases stat checks against things such as intelligence and Loyalty can cause a different response from the same hero to the same situation when the checks are repeated, althuogh I interpet it more as luck and/or having higher stats than actual learning.