---rww------------------------------------------------------------- 

statements

  (on B C)  (not (on B C))

actions

  (unstack B C)  (not (unstack B C))

  (question A1 A2 (does A2 (unstack B C)))

Will - belief, goal, intention, and action (no does)
> - Beliefs [e.g., (believes A (on B C)), (believes A (not (on C D)))]
> - Goals [e.g., (wants A (on B C)), (wants A (not (on C D)))]
> - Events [e.g., (does A (unstack B C))]
> - Intentions [e.g., (intends A (unstack B C))]

speach acts

   (inform A1 A2 (believes A1 (on B C)),
   (inform A1 A2 (intends A2 (unstack B C)))

   Will - (action A1 (inform A2 (belief A1 (on B C))))

   (intends A1 (does A2 (question A2 A3 (on B ?any))))
   (does A2 (inform A2 A1 (does A2 (question A2 A3 (on B ?any)))))

   Will X moved from Y1 to Y2
   (moved ?x ?y1 ?y2 ?t3) <--
      (at ?x ?y1 ?t1) (at ?x ?y2 ?t2) [(< ?t1 ?t2) (< ?t1 ?t3) (< ?t3 ?t2)]

---will------------------------------------------------------------

Just a few notes on Pat's email.

> First, we need to distinguish between conceptual relations that can
> hold in the world and actions that can alter it:
>
> - Basic relations [e.g., (on B C)] and negations [e.g., (not (on C D))]
> - Basic actions [e.g, (unstack B C)] and their negations

I don't think there's such a thing as a negated action unless the semantics are different than for beliefs, goals, and intentions. For example, (not (unstack B C)) may indicate an action that the agent attempted, but failed to complete. That seems like shorthand, but it might still be handy.

> - (question A1 A2 (does A2 (unstack B C)))

I would say that the modals for the memory partitions in an agent should be belief, goal, intention, and action (instead of "does").

> Speech acts are a special type of action that take a speaker as the
> first argument, a hearer as the second, and a belief, goal, event, or
> intention as the third. E.g.,
>

> - (inform A1 A2 (believes A1 (on B C)),
> - (inform A1 A2 (intends A2 (unstack B C)))

I think the jury's still out on the usefulness of speech acts as anything other than a labeling mechanism. I'm particularly concerned about treating them as special kinds of actions that are unlike any other kind of action at all. Is there a fundamental difference between Bob giving Sally his belief and Bob giving Sally his book?

More importantly, I think we can restate speech acts like this, which is more consistent with the above representation and could give us more flexibility. It doesn't confuse the verb "inform" with the modality of action.

(action A1 (inform A2 (belief A1 (on B C))))

> Relations and actions comprise the building blocks for ground content
> in working memory:
>
> - Beliefs [e.g., (believes A (on B C)), (believes A (not (on C D)))]
> - Goals [e.g., (wants A (on B C)), (wants A (not (on C D)))]
> - Events [e.g., (does A (unstack B C))]
> - Intentions [e.g., (intends A (unstack B C))]

"Event" doesn't seem like the right terminology. I'm not sure why you changed it here.

> - (intends A1 (does A2 (question A2 A3 (on B ?any))))
> - (does A2 (inform A2 A1 (does A2 (question A2 A3 (on B ?any)))))

See. Why do you have A2 repeated in the inform and the question? You only need the target, not the actor.

> - Conceptual knowledge takes the form of Prolog-like clauses with
>   a single consequent and multiple antecedents, some of which may
>   be negated.

In this representation, imagine the consequent as the name of a mental pattern. Any "before" and "after" relationships can be encoded in temporal constraints.

X moved from Y1 to Y2
(moved ?x ?y1 ?y2 ?t3) <--
  (at ?x ?y1 ?t1) (at ?x ?y2 ?t2) [(< ?t1 ?t2) (< ?t1 ?t3) (< ?t3 ?t2)]

Will

---pat------------------------------------


> Richard and Will - Here are some thoughts about what I'd like to see
> in the representation we're developing for reasoning about dialogues.
> I will start with the ground content

Richard and Will - I forgot to mention that each of the beliefs, goals,
events, and intentions will need time stamps to describe the intervals
during which they hold. We may be able to do this only for top-level
structures initially, but eventually we'll need to include them for even
deeply embedded structures. We can certainly say things like "Until
today, I believed you intended to do X before this weekend, but now
I believe you intended to wait until next month". More later, -Pat

-----
'
Richard and Will - Here are some thoughts about what I'd like to see
in the representation we're developing for reasoning about dialogues.
I will start with the ground content

First, we need to distinguish between conceptual relations that can
hold in the world and actions that can alter it:

- Basic relations [e.g., (on B C)] and negations [e.g., (not (on C D))]
- Basic actions [e.g, (unstack B C)] and their negations

I'm not certain whether we should include agents as the first argument
for actions, since they appear below.

Speech acts are a special type of action that take a speaker as the
first argument, a hearer as the second, and a belief, goal, event, or
intention as the third. E.g.,

- (inform A1 A2 (believes A1 (on B C)),
- (inform A1 A2 (intends A2 (unstack B C)))
- (question A1 A2 (does A2 (unstack B C)))

Relations and actions comprise the building blocks for ground content
in working memory:

- Beliefs [e.g., (believes A (on B C)), (believes A (not (on C D)))]
- Goals [e.g., (wants A (on B C)), (wants A (not (on C D)))]
- Events [e.g., (does A (unstack B C))]
- Intentions [e.g., (intends A (unstack B C))]

These structures can be embedded arbitrarily deeply to encode complex
beliefs, goals, events, or intentions. E.g.,

- (believes A1 (believes A2 (on B C))
- (wants A1 (believes A2 (on B C))
- (intends A1 (does A2 (question A2 A3 (on B ?any))))
- (does A2 (inform A2 A1 (does A2 (question A2 A3 (on B ?any)))))

Second, we need to encode generalized structures that match against
elements in working memory:

- Conceptual knowledge takes the form of Prolog-like clauses with
   a single consequent and multiple antecedents, some of which may
   be negated. The same consequent may appear in more than one clause.
   Together, they specify a conceptual hierarchy or lattice in which more
   specific predicates are above more general ones.

- Action knowledge takes the form of hierarchical Strips operators or
   a more distributed form with the same content. Each action clause
   (or distributed variant) should specifcy:

    - The action's name and arguments
    - The conditions under which the skill applies
    - For a nonprimitive action, its ordered subactions and arguments;
       for a primitive action, an executable Lisp function
    - The expected effects when applied under these conditions

We will also need some way to represent abstract relations between
the conditions and effects of actions, on one hand, and the beliefs and
goals against which they match, on the other. These appear necessary
to link the different steps into a narrative, whether at the domain
level or the dialogue level.

Best regards, -Pat