José Júlio Alferes, João Alexandre Leite, Luís Moniz Pereira, Halina Przymusinska, and Teodor Przymusinski

Dynamic Logic Programming.

 


Overview of interactions

N:o Question Answer(s) Continued discussion
1 12.12  Erik Sandewall
27.1  Luís Moniz Pereira
 

Q1. Erik Sandewall (12.12):

Your paper addresses update of knowledge bases represented by logic programs - a topic which is known and understood by only a limited part of the reasoning about actions community. Because of the similarity of research goals and the difference of background, I think that a discussion between you as authors and the readers in our area is particularly important and valuable; it can hopefully facilitate very much the understanding of this specific work as well as the "school" that you represent. Towards the end of the paper, you explicitly mention "reasoning about actions" as one of the intended applications, which of course is of particular interest to us.

Let me start out with a question about the prehistory of the approach you have chosen. You refer to Marianne Winslett's 1988 article [c-aaai-88-89] as an early reference for an "interpretation update" approach. Based on an example where it does not seem to give the intended result, you propose that the principle of inertia should be applied to the rules of the initial program rather than to the individual literals in a model. However, Winslett's article was written in response to an earlier article by Ginsberg and Smith [s-Brown-87-233] where they proposed exactly this: to define update on a set of logic formulae. Winslett pointed out some examples where the approach of Ginsberg and Smith did not work as intended, which is what led her to propose interpretation update. (An even earlier reference would of course be to the work by Lewis on counterfactuals [mb-Lewis-73].

My first question, therefore, is to what extent is there a difference: does your approach avoid the problems observed by Winslett, and if it does, what is the key to this improvement?

My second question is with respect to updates in the presence of observations and action laws. One of your results is that if the initial program is just a set of facts, then program updates and model updates coincide. However, in the case of reasoning about actions, one typically deals both with facts about the world at various points in time ("observations") and with rules characterizing some of the effects of actions ("action laws", "effect laws"). If update methods are used for characterizing ramification, which is what Winslett's article was all about, then presumably one wishes to prefer changes of "facts" (that is, sign reversal of literals) over changes of the action laws, at least as a first approximation. Only in the presence of accumulated evidence is it reasonable to revise a well established action law. How would you foresee representing such cases: will action laws be written out explicitly as logic-program rules, and what updates will then be obtained on the current state? On the other hand, if action laws are not represented as rules, how are they represented and how are the results in your article to be used?

A1. Luís Moniz Pereira (27.1):

It's probably easiest and less time consuming that I reply personally, rather than producing a reply involving all five authors, at least at this stage, where I'm not saying anything controversial, I believe.

  My first question, therefore, is to what extent is there a difference: does your approach avoid the problems observed by Winslett, and if it does, what is the key to this improvement?

Section 5.1 in our paper, and Theorem 7 in particular, shows that our approach generalizes that of Winslett as it applies to logic programs. Winslett updates interpretations instead of theories. We update theories or programs, but obtain interpretation updates as a special case. In the introduction of the paper we motivate why we do so. Further elaboration on the motivation can be found in our own work, references [Lei97] and [LP97], which I'll gladly provide for those interested, or can be gotten from our home pages.

Note that we're NOT performing belief revision. We're performing updates. We are not dealing with contradiction. New information overrides old information and that takes care of conflicts between the two. If new information is by itself contradictory we do not revise it. If old information is seen to be contradictory itself upon introduction of new information we don't revise it. We do no revision at all. Ginsburg and Winslett on the other hand are preoccupied with revision.

Coupling updates and revision, for when contradiction arises, is future work. It seems to me we can do it based on our previous work in LP revision, as can be found say in [mb-Alferes-96] and references therein, or my home page for more recent work on revision of logic programs, which I've done a lot of. But I should refrain to comment any more here on belief revision since it is not the subject of our paper. I do think that updates and revision should be developed separately.

  My second question is with respect to updates in the presence of observations and action laws. One of your results is that if the initial program is just a set of facts, then program updates and model updates coincide. However, in the case of reasoning about actions, one typically deals both with facts about the world at various points in time ("observations") and with rules characterizing some of the effects of actions ("action laws", "effect laws"). ... presumably one wishes to prefer changes of "facts" (that is, sign reversal of literals) over changes of the action laws, ... Only in the presence of accumulated evidence is it reasonable to revise a well established action law.

Again, we're not doing revision at all. If a fact updates a "law" by going against its conclusion then the fact wins! However, if you want to have persistent laws, you simply make them part of every update. Although we mention the application to reasoning about actions in the paper, we relegate it to future work. We are working on it at the moment. To do so we use updates that connect a previous state in the premises to a future state in the conclusion of the update, which type of updates we did not introduce in the paper. Since this discussion is about the paper, I will refrain from discussing that here further.

  How would you foresee representing such cases: will action laws be written out explicitly as logic-program rules, and what updates will then be obtained on the current state? On the other hand, if action laws are not represented as rules, how are they represented and how are the results in your article to be used?

This is ongoing work. Basically, actions are modelled by rules whose conclusions are post-conditions, but these only come into effect in the next updated state, ie the conclusions have the form of an internal self-update (not defined in the paper...), and the premises contain the pre-conditions and the action name. By performing an (external as opposed to internal) update of the action name the self-updating rule gets triggered. One of the post conditions deletes the action name if the action is not a persistent one. This will be the subject of a future paper.

I'm sorry to sound rather evasive, but the issues you address regarding actions are not tackled in the paper, but only mentioned as future work.

What the paper does is to establish a foundational base for dealing with updates of logic programs, which was sorely lacking till now. The exploration of a more dynamical type of logic programming is thus made possible. It will, we hope, lead to the use of present and forthcoming efficient LP implementational technology, and recent semantical developments, to a wider scope of problems.

References:

mb-Alferes-96José Júlio Alferes and Luís Moniz Pereira.
Reasoning with Logic Programs.
Springer Verlag, 1996.


Additional questions and comments, as well as the answers from the author(s), will be added into this structure. They will also circulated by email to the area subscribers. To contribute, please click Msg to moderator in the column to the right, send your question or comment as an E-mail message.

For additional details, please click Debate procedure there.

This debate is moderated by Erik Sandewall. The present protocol page was generated automatically from components containing the successive debate contributions.