Do we really want a good AI though? People love to bitch about bad AIs in games, but I bet even most codexers would ragequit when faced with opponents who play defensively, kite them, abuse chokepoints and line-of-sight shenanigans.
Yes. If better AI makes someone ragequit, that's not a bad thing. That's just separating wheat from chaff.
Those examples are not evolutionary game AI in any way, just a few sensible combat behaviors for classic state machine AI. Take Underrail for example. It has AI types with behaviors like that - human soldiers who attempt to break LoS after shooting at you, creatures that kite by spitting poisonous spikes at you and then keeping distance, other creatures that play defensively with a handful of behaviors (sneak up to victim, poison/paralyze it with their sting, retreat and hide, sneak back and attack helpless victim). The best is thing is that there's nothing new or special about any of these, they don't require endless resources or new tech to achieve. Basic game design and programming competence is enough.
Maybe the human intelligence of RPG devs and players alike has gone down the drain, but there are always glimmers of hope.
Yes, but... Dijkstra algorithm is the most cost effective search for the shortest path in a graph, but it is not a necessary basis. A* requires specified nodes and weighted edges, therefore "potential field" path planning are used when such a graph does not exist and cannot be made.
A* fails when all entities are able to move simultaneously as you can see in Pillars of Eternity when a paladin is running around enemies in circles instead of attacking closest.
Are you guys current computer science students by any chance? Nobody else would nitpick algorithm details... The point was that game AI hasn't seen any significant improvements after home computers became powerful enough for real-time pathfinding that doesn't suck (see early RTS games for examples that suck), and A* is the industry standard for that. Implementation details vary.
However, there have been some smaller improvements. Probably fairly large improvements from gamedev perspective.
Back in 2005, FEAR did something new. Basically it added another layer of abstraction on top of the industry standard finite state machine AI to make implementation of complex NPC AIs faster & easier. But honestly, from player perspective, FEAR's AI (and all newer games using similar systems) don't do anything particularly impressive compared to older A* and FSM systems. FEAR also had a simple squad controller on top of the individual AIs, but there weren't any complex behaviors programmed in it - any seemingly complex squad behavior was just emergent happenstance. NPC maneuvered to flank you? Only because it couldn't find valid cover/route elsewhere. And they hyped that system to death...
STALKER was another game with overhyped AI system. Its ambitious ALife system was completely neutered before anything was released to the public. Most of the games' AI behaviors come from smart terrains that give NPCs tasks. Unlike marketing/hype claimed, AI stalkers can't live and progress through the zone like the player does. At all. They simply get assigned tasks by these smart terrains - campfires, faction bases, sleeping bags, etc. And at some point during early development STALKER SHOC was pushed towards more scripted storyline, less sandbox. My understanding is that this meddling was publisher demands, but don't quote me on that. So they had to implement space restrictors where the AI couldn't go and do its thing before the player has progressed through the storyline in that area. IIRC the only special thing the ALife system ended up doing was its AI LOD system - it allowed real-time simulation of all NPCs in the entire zone by switching far-away characters to simpler simulation mode. I think it was the first game with complex level geometry to do this, but older X universe games had similar AI LOD system allowing real-time simulation of far larger number of AI actors.