I'm going to be using one of the common game AI techniques, such as
some form of finite state machine or perhaps one of the slightly newer
approaches such as motivational graphs or fuzzy cognitive maps. My
current "favourite" would be a subsumption architecture, which has each
character being driven by multiple, concurrently active FSMs, with the
lowest level one controlling low level behaviour (eg. taking a step
forward, jumping, etc...) and higher ones covering increasingly complex
behaviour (eg. walk to the table, buy a drink, etc...). My concern is
that having such a system in place might prove a bit too expensive if
you are dealing with hundreds of characters at a time. I'll have to
investigate optimisation techniques, the balance of behavioural
complexity, number of characters and performance, and so on. I also
plan on having controls that let the player (not that it is going to be
a "game") turn on or off various behavioural traits (eg. rules or
states in the character behaviour) to see the effect they have on the
believability of the scene, and the performance.
I won't be having any sort of learning; given that I'm researching AI
for BACKGROUND characters in games, the resources that are
realistically going to be allocated to it in a game (in terms of memory
and CPU time) are very low. Learning would be almost entirely wasted on
characters that serve no real purpose other than to make an environment
more believable.
[img]http://card.mygamercard.net/sig/Default/reno84.png[/img]
Designer @
Haiku Interactive |
ReNo-vation.net