Erik SandewallLogic-Based Modelling of Goal-Directed Behavior. |
[full text] |
[msg to moderator] [debate procedure] [copyright] |
N:o | Question | Answer(s) | Continued discussion |
---|---|---|---|
1 |
13.4 Rob Miller |
13.4 Erik Sandewall |
|
2 |
Anonymous referee 1 |
17.7 Erik Sandewall |
|
3 |
Anonymous referee 2 |
17.7 Erik Sandewall |
|
4 |
Anonymous referee 3 |
17.7 Erik Sandewall |
19.8 Anonymous referee 3 19.8 Erik Sandewall |
Additional reviewing information: Minor details noticed by reviewers.
Q1. Rob Miller (13.4):
Dear Erik,
I'm confused about axioms S4, S5 and S8 involving the action
S4 seems to state that it is always possible to perform a test to see if a given property is true. But I would have thought that the opposite is nearly always the case - that in most commonsense or robot environments we could almost assume by default that the agent is not in a position to directly test the truth of a given property of the environment. (This is certainly true for me as I type this question at home. I can't check the condition of anything in my office, in the other rooms in my house, etc. without first performing some other actions.) So what is the rationale behind axiom S4? Is the agent perhaps "testing" it's own knowledge, rather than the real environment?
S5 states that test actions are always instantaneous. Why should this be so, at least in a framework where other actions are not instantaneous?
S8 says that, once executed, a test action fails if and only if the corresponding property is false. But this seems to go against the commonsense notion of "testing". For example, when I test to see if I have any new emails (actually by clicking on the "?" button on Netscape) I count this action as successful if the system reports "no new messages on server". I count the action as failed if (as frequently happens with my faulty modem here at home), clicking on "?" crashes my computer. I'm then left with no information as to whether I have new emails or not. So is this not a test action in the sense that you intended?
Regards,
Rob
A1. Erik Sandewall (13.4):
Dear Rob,
The "action"
Will
The axioms assume that the argument of
The mnemonic "test" was inspired by an opcode in an assembly language that I learned a long time ago.
Regards, Erik
Q2. Anonymous referee 1:
Generally, the paper has the flavor of a progress report rather than a well-knit article. My understanding is that this perfectly complies with the publication policy of ETAI. However, sometimes such a status of a paper is indicated by adding the qualification "preliminary report" to the title, which the author might consider in the present case as well. The author himself admits that there is still a lot of work to be done. In particular, a number of rather complex axioms are proposed, of which it is claimed that they characterize "obvious properties" (p.5) of relation symbols which have been introduced with a certain meaning in mind. Past experience taught us that it can be dangerous to propose axiomatizations which seem most natural at first glance. I therefore strongly support the author's remark on the need for an "underlying semantics, and a validation of the present approach with respect to that semantics..." (p.16). An essential, most basic property of any useful axiomatization is consistency. Maybe the author could add a result in that direction.
p.7: The purpose of Section 2.6 is not clear. Composite fluents aren't used elsewhere in the paper, and so one wonders, for instance, why logically equivalent composite fluents cannot be treated as equal.
p.8: The operators
p.8: The predicate
A2. Erik Sandewall (17.7):
Generally, the paper has the flavor of a progress report rather than
a well-knit article. My understanding is that this perfectly complies
with the publication policy of ETAI. However, sometimes such a
status of a paper is indicated by adding the qualification "preliminary
report" to the title, which the author might consider in the present case
as well.
|
The author himself admits that there is still a lot of work to
be done. In particular, a number of rather complex axioms are proposed,
of which it is claimed that they characterize "obvious properties" (p.5)
of relation symbols which have been introduced with a certain meaning in
mind. Past experience taught us that it can be dangerous to propose
axiomatizations which seem most natural at first glance. I therefore strongly
support the author's remark on the need for an "underlying semantics, and
a validation of the present approach with respect to that semantics..."
(p.16).
|
An essential, most basic property of any useful axiomatization is
consistency. Maybe the author could add a result in that direction.
|
p.7: The purpose of Section 2.6 is not clear. Composite fluents aren't
used elsewhere in the paper, and so one wonders, for instance, why
logically equivalent composite fluents cannot be treated as equal.
|
p.8: The operators
|
p.8: The predicate
|
Q3. Anonymous referee 2:
This is a well-motivated and timely contribution to the field, and, perhaps with some presentational improvements, should be accepted for publication in ETAI. The paper's main original contribution is an axiom-based account of control, focussing on the issue of action failure, retry and replanning. It also addresses the issue of how these axioms can be integrated with logical theories of action.
There follow a number of general questions I would like to raise, some presentational suggestions, and a list of typos.
Although I'm attracted to the author's work, I am tempted to ask why we should want an axiomatic account of goal-directed behaviour. For example, Axioms G1 to G4 really constitute an algorithm, and perhaps make better sense presented as an algorithm. From my point-of-view, the role of logic in cognitive robotics is to represent the world, not to represent the internal workings of the agent. Perhaps the author would like to address this question, in an electronic discussion and/or in the final version of the paper.
In Section 2.5, the third formula on page 7 entails that the
postcondition of an action is false until the action is finished. But
consider the example of a compound action
It's not clear to me what it means for the robot to "state" a formula on page 9. Does this mean it passes that formula to a control module?
I have one serious presentational objection to the paper. I think it would be very much improved if there were some concrete examples showing what inferences can be drawn from the axioms. (By a concrete example, I mean one with meaningful fluent names.) Even a single concrete example illustrating some of the main ideas would be of enormous benefit to the reader's intuition. Otherwise, a hostile reader might think this is all just empty formalism.
Here are some more minor presentational suggestions. At the start of
Section 2.6, the author introduces logical connectives for composing
fluents. But these have already been used in Section 2.5. Perhaps this
material could be reshuffled. Similarly, on page 10, the author declares
how the variable symbols
A3. Erik Sandewall (17.7):
Although I'm attracted to the author's work, I am tempted to ask why we
should want an axiomatic account of goal-directed behaviour. For
example, Axioms G1 to G4 really constitute an algorithm, and perhaps
make better sense presented as an algorithm. From my point-of-view, the
role of logic in cognitive robotics is to represent the world, not to
represent the internal workings of the agent. Perhaps the author would
like to address this question, in an electronic discussion and/or in the
final version of the paper.
|
Furthermore, if this software is described in a purely algorithmic style, it is difficult to avoid introducing a lot of administrative machinery. This is exemplified by the article "A Formal Specification of dMARS" by Mark d'Inverno et al. (Proceedings of the ATAL-97 workshop, pages 155-176), where a variant of PRS is specified in the Z notation (which is a software specification language). In my approach, the agent's rational behavior is specified in terms of restrictions: for each scenario, the set of models for the formalization is intended to equal the set of rational behaviors, each interpretation being one possible history in the world. Such a constraint-oriented approach has chances of being much more concise and to the point.
Finally, once we decide that it is appropriate to specify rational robotic behavior in logic, we also have to deal with actions. The robotic behavior consists of actions, so this already defines a connection point. Also, when several agents are involved, the actions of one need to be observed and understood by the others. In the longer perspective, I think we will see a lot more integration between theories for reasoning about actions and for characterizing agent behavior.
In Section 2.5, the third formula on page 7 entails that the
postcondition of an action is false until the action is finished. But
consider the example of a compound action
|
As another simple example, consider the case of actions which are
described in terms of a precondition, a prevail condition, and a
postcondition, where the postcondition is at the same time the
termination condition for the action...
|
It's not clear to me what it means for the robot to "state" a formula on
page 9. Does this mean it passes that formula to a control module?
|
I have one serious presentational objection to the paper. I think it
would be very much improved if there were some concrete examples showing
what inferences can be drawn from the axioms. (By a concrete example, I
mean one with meaningful fluent names.) Even a single concrete example
illustrating some of the main ideas would be of enormous benefit to the
reader's intuition. Otherwise, a hostile reader might think this is all
just empty formalism.
|
Here are some more minor presentational suggestions. At the start of
Section 2.6, the author introduces logical connectives for composing
fluents. But these have already been used in Section 2.5. Perhaps this
material could be reshuffled. Similarly, on page 10, the author declares
how the variable symbols
|
Q4. Anonymous referee 3:
The paper relates the concept of a "deliberated retry" to the author's earlier work on the logic of actions and change. It is intended for use in an applied project related to controlling an intelligent airborne vehicle. The problems discussed in the paper are important, and its ideas are interesting and original.
On the negative side, the paper does not really prove that its formalism is good for any specific purpose. It does not even make any mathematically precise claim of this kind. It would be good to include a description of goal-directed behavior in a toy domain and prove that some intuitively expected conclusions abour goal-directedness do indeed follow from the given axioms using the entailment methods proposed by the author. This would be more convincing than the "hand-waving" arguments in favor of the proposed approach given in Sec. 4. In the absence of such an example, the paper is reminiscent of the work on actions done in the early days of AI, such as the "frame default" in Reiter's 1980 paper on default logic, or the formalization of the blocks world in McCarthy's 1986 paper on circumscription. The ideas were interesting, but their authors were unable to prove anything about them. As the familiar shooting scenario demonstrated, a nonmonotonic formalization that looks plausible may turn out to be unsatisfactory after all. If the author of this paper tries to check that his theory works for one or two toy examples, he may very well discover bugs that need to be fixed.
It seems to me that Rob Miller was right when he suggested in his message
to the author that
Re Theorem 1: There seems to be an implicit assumption here that the set of intervals in question is finite. Shouldn't it be included in the statement of the theorem?
Re Theorem 2: I am puzzled by the use of the word "conversely" here. It seems to me that both parts say more or less the same thing.
In the 3rd displayed formula on p. 7, conjunction is applied to fluents,
which is only explained in the next section, and an interval is used as
the first argument of
My recommendation is that the paper be accepted for publication in the ETAI after the author makes the changes that he deems appropriate.
A4. Erik Sandewall (17.7):
On the negative side, the paper does not really prove that its formalism
is good for any specific purpose. It does not even make any mathematically
precise claim of this kind. It would be good to include a description of
goal-directed behavior in a toy domain and prove that some intuitively
expected conclusions abour goal-directedness do indeed follow from the
given axioms using the entailment methods proposed by the author. This
would be more convincing than the "hand-waving" arguments in favor of
the proposed approach given in Sec. 4.
|
You might say that the paper ought not to be published until such a validation exists and can be included. This would be an excellent position to take, provided that it were shared by the present community. However, even a quick look at the literature in our field will show that that's not the way things work: Citations to earlier work refer almost exclusively to the approaches proposed by the earlier author, and quite rarely to the theorems that were proved in the earlier paper.
Besides the formal validation, it's the applications that will prove what the formalism is good for, but this would also take too long in the present paper.
In the absence of such an
example, the paper is reminiscent of the work on actions done in the early
days of AI, such as the "frame default" in Reiter's 1980 paper on default
logic, or the formalization of the blocks world in McCarthy's 1986 paper
on circumscription. The ideas were interesting, but their authors were
unable to prove anything about them. As the familiar shooting scenario
demonstrated, a nonmonotonic formalization that looks plausible may turn
out to be unsatisfactory after all. If the author of this paper tries to
check that his theory works for one or two toy examples, he may very well
discover bugs that need to be fixed.
|
Another comparison with the work on the frame problem in the 1980's is more appropriate: those were the days of first exploration of the problem (besides for a small number of articles in the 1970's), and it is quite natural and respectable that one started by looking for approaches that were at least intuitively plausible. After that start it was natural to look for precise and proven properties of those approaches. Ramification and causality have followed suit, and qualification is in its early stages. The characterization of rational robotic behavior is just in its very first stage of development, that is all.
It seems to me that Rob Miller was right when he suggested in his message
to the author that
|
The present paper works with the simplifying assumption that the agent has perfect knowledge of the fluents. This is in particular in the interest of the reader, since it makes for heavy reading to address all problems at once.
Regarding the author's suggestion that the
incompleteness of knowledge be represented by distinguishing between
|
Re Theorem 1: There seems to be an implicit assumption here that the
set of intervals in question is finite. Shouldn't it be included in
the statement of the theorem?
|
Re Theorem 2: I am puzzled by the use of the word "conversely" here. It
seems to me that both parts say more or less the same thing.
|
In the 3rd displayed formula on p. 7, conjunction is applied to fluents,
which is only explained in the next section, and an interval is used as
the first argument of H which, as far as I can see, is not defined at all.
|
C4-1. Anonymous referee 3 (19.8):
Thank you for your reply to my comments. One point is still not clear to me. You claim that, in Theorem 1, there is no need to assume that the set of intervals is finite. The following seems to be a counterexample:
si = 1/(2*i+1), ti = 1/(2*i) |
C4-2. Erik Sandewall (19.8):
Theorem 1 states that under the axioms and for some ordering of the
intervals,
s-i = 1/(2*i+1), t-i = 1/(2*i) |
Your example also shows that if there is an infinite sequence of
intervals, the theorem does not hold in the limit as
Your example does however remind me that if one is going to be very
technical, it might be better to phrase the theorem so that it states
that if interval
13-Jun-98 20:28