Since back when my basic code was used for playerbot warrior I had an idea to make a very advanced pvp AI bot that learns from experience. I'm still working on it and speaking to some psychology professors from my university to see what best ways for pvp bot to learn are, as well as thought processes and others.
As far as my work is going right now, the pvp bot identifies "events" as either intervals of time, or scenario triggers (battle, death, killing blow, identify enemy(s)), and inserts them in seperate db i created in sd2 in table "ai_memory", with columns including LOTS of factors that define the "event", along with some kind of indication of whether a positive or negative outcome with the decision made. Based on searching through db from previous events (or after thinking of virtual memory and what happened from pvp bot decisions), the bot takes multiple params of his surroundings, hp, hp ratios, allies, etc etc. and makes thought process to see what to do (fight, retreat, support etc) based on all params.
Right now I'm working on trying to create a "mood" or personality almost that will alter decision making, creating flaws of being coward, or too cocky trying to rambo into a 1v5 battle after being on massive killing spree... and its annoying >.>
I've been researching AI for a bit and thought what better "test zone" than mangos. Only thing is this is not controlled via player, the ai will have a "mind of its own" almost. However its pathing is something I've yet to even think about, currently what i've got it doing is just mindlessly roaming around set waypoints within warsong and so far recording some information on its surroundings.
Maybe though it can be implemented in your system as soon as I have some working copy able to release