OK so this is a generalized simulation for many of the entities (exact bookkeeping isnt needed).
Long ago I looked into this same idea and split entities between :
"Dum' types (local flavoring, move not so much from their area, wake up/auto-generate when player is nearby -- non-persistant)
These probably won't be doing long range pathfinding, and even then only enough medium movement range to appear to be acting naturally.
You can still have them do ecosystem interactions - locally (predators hunting prey, herbivores moving between grazing and water, etc..)
Big map migrations and such is handled by adjusting patterns on high level coefficients which then control the spawning content.
For complex simulation patterns, high level control entities on the 'big map' could be persistant, they effect/control shifts in coefficients of areas (cellular atomata methods).
Pathfinding for these is generally local (still cross chunks) but depending on how realistic and detailed you want them, many might run on schedule driven repetative medium level paths (compute set once, and then local A* immediate steps including dynamic (if you have blocking by other entities).
Unfortunately, dense terrain is a favorite environment for many animals BECAUSE its difficult for their predators OR its where alot of thesustinance is located. But then dense also greatly slows movement, so more real time to decide on any longer pathing)
The high level (persistant) entities will navigate around major blocking obstacles ocean/rivers/mountains deserts and will set direction for long distance migration of herds of animals. High level simulation can be done (group fo predator interacting with group/herds that wander by, etc..) These high level entities still can have migration patterns - precalc'd general paths. When the player comes close, then whatever interactive entities are realized (auto generated). and have general behavior motives set to do their appropriate behavior variants.
--
'Smart' (persistant, entities that have very complex player interactive/motive driven which will/can follow players around the big map with intent, requiring more exact bookkeeping because players see them over and over and result of interactions would be important and persistant)
These, even though they can have 'generalization' when well away from where the player sees them, have to transition to full detail sufficiently far away to be able to 'Act Natural' by the time the player sees them (which can include forcing high detail around where THEY are centered to a sufficient area to all the things THEY interact with). Their long range pathing has to be more exact/realistic/proper so they wont be where they shouldnt be (and are going about their proper 'business') when the player comes near and they are visible.
These entities (probably other humans) have the far more complex AI (and interacting with them is a good(more interesting) part of the game).
You can STILL auto-generate them (like if the game is a life journey and you cross the world and probably will never meet those local 'smart' residents again). You use Templates (hierachical parameterized templates) to generate these local occupants (probably with a largely hand built seed map) and likewise use the high level persistant entity (tribes?) to make their overall simulation flowing/balancing across time. Those Templates and the rules that direct them are the major difficulty (work to get them to be plausible) for the auto generation when used to generate secondary characters ('main' characters probably will be mostly hand crafted)
These can do the long ranged pathfinding (but still probably will have daily/weakly/seasonal reused paths) - alot more data, but there are many fewer of them compared to the 'Dum' background entities. You probably will find that their pathfinding is the LEAST part of their AI processing (you might use things like Planner driven AI which use pathfinding as a mere 'tool')
-
It all depends about how detailed the game is to be (and how much mundane stuff is cut out)
Funny thing is in one design (for a very detailed game) I actually had a 'micro' simulation for crafting activities (you kill an animal and now you have to butcher it and convert it for transport, or 'gathering' of edible/useful plants in a small area (or even just gathering firewood). Game where (if you want to look "there is something interesting under every rock"). Using the same methods to LOD a small area into very fine details .... Thats usually overkill for a 'game' and at that point it was more 'simulation'.
Likewise it is the transitioning -- close high detail with a boundry shifting into the lower detail 'generalized' simulation, which is much of the programming headache.
---
Another thing - once you do the autogeneration to fine detail (can be alot of processing if you have intricate Templates) you dont 'throw it away when the player walks on (he might suddenly decide to turn around and go back, or constantly move around the same large area. So I had a system to save the detailed chunks for an extended time (the 'Dums' were local on them) - roll out to disk, and then 'roll them in' when the player comes back in range and 'patch' the local details to compensate for any intervening time (and have any persistant 'smart' entities added back to them). If you are gone long enough any changes the player made can melt away and you can throw away that saved local chunk. But the autogeneration would have to be sufficiently deterministic for much detail for when the player later comes back again (particularly if your terrain isnt overly generic' and specific terrain matters).
My chunk data was actually encapsulated (a block of map memory with its own local dynamic object heap, internal offset pointers) so that a simple memory copy of the whole thing could save/restore it all with minimal processing. They would constantly get loaded ahead of the players visibility (partial low detail LOD out furthest away). Again depends on your game's detail relevancy if you have to go that far with such a mechanism.
--------------------------------------------[size="1"]Ratings are Opinion, not Fact