Since this is a position statement, I suppose it's legitimate to give a
sales pitch for a particular formalism, and to be a bit controversial.
AI needs an action formalism that is expressive, and that incorporates a
solution to the frame problem that's robust in the face of the phenomena
it can represent. The formalism should be expressive enough to represent
at least the following phenomena.
Actions with indirect effects (ramifications)
Concurrent action
Non-deterministic action
Narrative time
Continuous change
A rigorous argument that the formalism in question solves the frame
problem should be supplied.
Here comes the controversial bit. I believe that the most instructive
way to validate a formalism is through a judiciously chosen set of
representative benchmark scenarios. I sometimes feel that attempts to do
this validation by proving a relationship between a formalism and some
other formal structure are just an excuse for introducing a lot of
unnecessary mathematics. I strongly believe that we can only contribute
to AI if we divert our efforts away from proving "interesting" theorems
and into the application of our formalisms to fundamental problems in AI
(such as planning in robots).
Here comes the sales pitch. The event calculus presented in my
forthcoming book is capable of representing all the above mentioned
phenomena, and the solution to the frame problem that accompanies it is
provably immune to the Hanks-McDermott problem. In recent ECAI and AAAI
papers, this formalism is applied to a serious example involving a
robot. The example includes all five of the above listed
representational features. If your favourite action formalism can't do
this robot example by now, maybe it's time to change to a new one.
Honesty time now. The event calculus cannot, as yet, handle knowledge
producing actions, or complex actions (ie: actions including program
constructs). To me these requirements seemed less urgent than the ones
in my list, but they're undoubtedly important.
Finally, the last thing we need is new formalisms. The only excuse for
introducing a new formalism is that it is fundamentally different from
any of the existing ones. Instead, we have to try to expose the
underlying similarities and differences between possible formalisms in
the hope that we can start to understand the range of representational
choices in a principled way.
The above opinions are those of the author on 20th November 1996, and
may be subjected to total revision when he is older and wiser, or
possibly sooner.