Overview of interactions

Q1. Workshop discussion:

There were a number of comments by several participants on the question of chronological minimization vs chronological ignorance, and the author's reference to Shoham's work. The bottom line seemed to be that we have to make distinction between chronological minimization of change which was proposed by several authors concurrently in the flurry of the Yale shooting problem, and chronological minimization of ignorance which was developed by Yoav Shoham in his thesis and subsequent book. The early formulations of chron.min. of change were by Shoham (conference article), Kautz, and Lifschitz.

Several authors have investigated chron.min. of change in more recent work, whereas chron.min. of ignorance has not been further used. There are also indications that the latter gives incorrect results for many simple examples. However, the present article does seem to be using chron.min. of ignorance.

Q2. Erik Sandewall:

I have objections to some parts of your argumentation at the beginning of the article, and finally some questions.

In the article, you write that I have "attacked" Shoham's causal theories for technical reasons, and that I "went on to propose increasingly more complex forms of chronological minimization", with reference to my 1994 book. This is inaccurate, since my approach in [mb-Sandewall-94] as well as elsewhere has been to compare and analyse different entailment methods, rather than advocating one of them over another. In particular, the book contains assessments for about one dozen such methods, several of which have nothing to do with chronological minimization. The set includes both a number of methods that had previously been introduced by others, and methods that naturally arose in the course of the work.

In the article, you also object to the use of filtering, where observations are imposed after the minimization of changes (or of discontinuities) has occurred. This method is now fairly widely used; it was adopted by Lifschitz under the name of nested circumscription. Your objection is that it is complex and that it does not have any counterpart in everyday causal reasoning. However, if simplicity is an issue, then I can offer it through PMON, also defined in [mb-Sandewall-94], which is about as simple as you can get.

With respect to having a counterpart in everyday causal reasoning, this introduces an issue that the book makes no attempt and no claim to address. The criticism is therefore irrelevant. The analyses of chron.min. as well as of PMON in the book only consider their use for the plain inertia problem; it is nowhere claimed that the analysis would apply for any kind of causal reasoning.

I am afraid that at this point you share in a habit that is quite common in our field, namely to criticize an approach X only because it fails to handle some example correctly, and regardless of whether it was claimed or intended to handle that case. In other words, you can be criticized for not solving all the problems at once. This style of discussion may be understandable in those cases where the author has not specified what class of problems his approach is claimed to solve in the first place, of if the author has made excessive claims. However, since I have tried to make a point of specifying what topics I am addressing at each point, I find your objection misdirected.

With respect to "everyday reasoning", please see my position statement (forthcoming) in the panel discussion.

These were my objections to your argumentation. Finally, a few questions:

1. In view of your reference to "everyday causal reasoning" in the context of my work, is the logic proposed in your article intended to match such reasoning with respect to both the structure of the inference process and the conclusions obtained? In particular, what sense does it then make to treat prediction and explanation as two different reasoning processes, as you propose to? After all, in everyday reasoning we are able to combine the two seamlessly.

2. In your article, you introduce the predicates Aff and Qual to be used in representing inertia and change. How are these related to similar predicates previously proposed by others? The Aff predicate seems to be occlusion, once again reinvented, and your general inertia rule seems to be the same as the nochange axiom schema used in PMON (originally introduced by McCarthy I believe).

3. The most important difference between your approach and what has been proposed before appears to be that you are using a threevalued logic; is this correct? If so, and in view of the similarity of other aspects of the approach to previous work, what advantage do you derive from switching to a three-valued logic for this purpose?

References:

This on-line debate page is part of a discussion at recent workshop; similar pages are set up for each of the workshop articles. The discussion is organized by the area Reasoning about Actions and Change within the Electronic Transactions on Artificial Intelligence (ETAI).

To contribute, please click [mail to moderator] above and send your question or comment as an E-mail message.