I strongly believe that if you ever want to model any type of converstion reasonably, you've got to model the knowledge behind it. And I'm not even talking about fluid natural 'talk about anything' conversation. I'm talking about some rudimentary ability to converse with a character and feel like you're actually communicating to the character.

I think that's important. That the player gets the feeling he's actually communicating with the character.

I'm working on what I would call a 'common-sense' agent. The agent has a set of axioms that give it common sense about things in life that are necessary to know to get by. These are basic principles that are common to most (not all) cultures both past, present, and future. The idea is to then embellish this with specific domain facts related to the game-world and the life of the character.

Some of the common-sense theories I'm working on are:
-----------------------------------------------------
spatial relation
ownership / possession
family
social relationships
social manners
shelter / clothing
food
money management
time
scheduling
conversation
questioning
deliberate remembering
sleep
work skills
geography / politics
crime and punishment
death / survival
hierarchy / classification of things
transportation modes
teamwork / hiring
color / aesthetics
lying / misinformation
weather
land / forest / mountains / ocean

Note that when I say something like "geography / politics" I'm not talking about any specific nation or government, but the simpler idea of what it is all about. Something like "Nations have leaders and boundries and capital cities".




Edited by - bishop_pass on November 24, 2000 2:23:43 AM

I''d agree that some representation of knowledge would be necessary.

But I''m not convinced that the NPCs need logical deduction.

So the axioms that you list would be the general categories of knowledge?

JSwing,
The categories I list above are some or most of what I think would help for an agent that can behave autonomously or independently or somewhat rational anyway.

For example, under the ownership category, all agents would have axioms like the following:

If (x owns y) &
(z subcomponentOf y)
->
(x owns z)

That means: If you own something and something else is a sub-part of that something, then you also own that sub-part too.

Another ownership axiom might be:

If (x sells y) &
(z buys y)
->
(z owns y) &
not (x owns y)

I think I see. The axioms are a chain of deductive reasoning that can be used for any appropriate objects.

I still politely disagree that deductive reasoning is required for conversation. Items that are essentially fixed (backstory) cand be hardcoded since they won''t change. Items that are dynamic can alter the knowledge of any observers as they change. The deductive axioms are replaced by a set of transformations to the NPC''s associative set of nodes.

If I was going to use the knowledge for something larger than conversation, then the deductive reasoning would be vital. I''m not planning anything that big (yet).