Wombah, BrianL: STRIPS does indeed have many limitations (as you rightly point out, it's very old now!), I was just pointing it out as it helped me learn a lot about planning in general. There have been many proposed modifications that make it more useful, however. Also, it may well be the case that it's only a total order planner, I was going purely on lecture slides from my course at Warwick Uni; given the poor quality of other modules it's entirely possible the information is wrong! I'll have to double check my sources next time :)

While Markov Chains are 'nice' for solving some problems, I really don't see how they can be complex enough for convincingly good AI (i.e. better than current technology). That, and they generally solve different problems to planners. I've been intending to design a general purpose AI architecture for some time now, and it's my opinion that a combination of many techniques is needed. A planner is great, as long as the agent knows its goals; some utility based decision function needs to be included. Some element of uncertainty needs to be accounted for too, perhaps even basing the framework around fuzzy logic. Learning is very important too, if not in real time then for training AI players to act like humans before release of the game. No single feature will make the AI realistic, only a well-chosen combination of features.

Quote: Original post by hymerman
While Markov Chains are 'nice' for solving some problems

Markov Chains are a tool for representing processes containing uncertainty (be they actual processes or constructed ones). In planning domains, Markov Decision Processes - and more generally Partially Observable Markov Decision Processes - are representations of planning problems, for which solutions produce a plan of action (a sequence of decisions that will achieve a goal).

Quote: I really don't see how they can be complex enough for convincingly good AI (i.e. better than current technology).

No one in AI would expect a Markov Chain to be a good AI. However, very good AIs have been created around POMDPS.

Quote: Some element of uncertainty needs to be accounted for too, perhaps even basing the framework around fuzzy logic.

I can't help but cringe when I read sentences like that. Fuzzy Logic is NOT a tool for representing uncertainty. It fails several basic mathematical requirements for such. It is a tool for non-binary set membership descriptions and may be used for a variety of purposes. The closest to uncertainty modelling would be modelling vagueness; that is, the varying degree to which something may belong to several non-distinct classes, which are usually discrete labels given to a continuous domain. To try and use FL for uncertainty representations is to misuse the tool and to fall short of the potential of a true uncertainty method.

Cheers,

Timkin

Okay, I officially quit this university. There I was mistakenly thinking it would teach me correct knowledge. Thanks for the corrections, anyway.