Joe, one of the key points in my paper with Judea is that causality is very much model-dependent. If you construct a "bad" structural model, you get "bad" answers about causality. You may be right that, for some applications, constructing an appropriate "good" causal model may turn out to be hard. A particuarly important issue is the choice of random variables. If you leave out the random variable for kicking the board, you may have a bad caussal model. Unfortunately, we have rather little to say about how you validate a model (in particular, how you can validate that you've chosen the "right" variables).
Interdisciplinary work may well help us here. There is work going on in the psychology (by people like Steve Sloman and Dave Lagnado) testing the extent to which our models predict the answers given by people. To be honest, the results are mixed. It seems very hard to describe what people do just by our models (although they do seem to give some insight). I'm sure much more work could be done along these lines. -- Joe >From [EMAIL PROTECTED] Mon Jul 17 09:05:43 2006 X-MimeOLE: Produced By Microsoft Exchange V6.5.7226.0 Content-class: urn:content-classes:message MIME-Version: 1.0 Subject: Structural models are necessary but may be impossible to validate Date: Mon, 17 Jul 2006 09:04:50 -0400 X-MS-Has-Attach: X-MS-TNEF-Correlator: Thread-Topic: Structural models are necessary but may be impossible to validate Thread-Index: AcaoIQApez8As2AdQ2WAieQ3qkAfWwBez1Zw From: "Mitola III, Joseph" <[EMAIL PROTECTED]> To: "Joseph Halpern" <[EMAIL PROTECTED]>, <uai@engr.orst.edu> Cc: <[EMAIL PROTECTED]> X-OriginalArrivalTime: 17 Jul 2006 13:04:51.0817 (UTC) FILETIME=[97330590:01C6A9A1] Content-Transfer-Encoding: 8bit X-MIME-Autoconverted: from quoted-printable to 8bit by sundial.cs.cornell.edu id k6HD5gv20523 Professor Halpern, I'm concerned that the structural model building you appropriately suggest is something like weather reporting: it is easy to get the first and second order equations right, but the results in any given case depend strongly on the initial conditions that in turn have "critical" features that occur on such a fine scale that we had to wait for computing to improve by 1000x before being able to predict the weather as well as we do today, e.g. to predict hurricane tracks, good but still not perfect. This note suggests interdisciplinary research to address Zadeh's examples with your methods in a way that can be validated. Social systems (and Professor Zadeh's examples all entail social systems), unlike weather, have many parameters that seem to be unmeasurable no matter how fine the grid, depending in inscrutable ways on internal psychological processes of the decision makers (e.g. the buying public, the stock market, and a person under extreme stress in Zadeh's clever progression). Although one can retrospectively model such phenomena from the perspective in Pearl's superb mathematical treatment, the questions of observability of the underlying phenomena seem to limit one's ability to validate models so that they could be used not just to describe but also to predict (engineering predictions, at least about stability versus likely instability, sufficiently for decision-making insights, not fortune telling of social trajectories, Asimov's Foundation not withstanding). The first thing I learned about random processes at Johns Hopkins was that a probability space must be built on a measurable space with at least a sigma-algebra defined on it. The human psyche may be described as in some ways a random space, but so far, I haven't seen problem of measure, nor of aggregation of measures (the sigma-algebra) defined for the psyche. In decision support, one attributes value to game positions but doesn't include the probability that the other player will kick the board over if the losses are too big. This isn't funny or irrelevant but really may be a missing aspect of mathematical causality: the regular lack of measurability in Professor Zadeh's examples, thus the lack of a validated sigma algebra (arithmetic on 1/k is easy; validate the model with real world human aspects such as, e.g. voter fraud, is harder, so what is the sociological or psychological basis for "critical" and for 1/k?). So meaningful mathematical statistics or random process models work well in automatic control where we can isolate the aircraft meaningfully from the rest of the world except the air flowing over it, we can define a measurable space of air pressure, pistons, valves, and electronics, and thus we can define a fixed point corresponding to level flight and we can write the algorithms to sustain that fixed point in a validated way. Not so in the social sciences, like business and law: we don't seem to have the basic measurability, the basic algebra of measures that we can validate and agree to and thus the practicality of using the insights in Pearl's book seems to be hard to come by, maybe impossible in the near and mid term. But I wonder if this topic might not be a good realm for interdisciplinary research that tries to quantify the significant aspects of at least the sociology (not the psychology) of causality in business and law so that we can be on more solid ground than anecdotal evidence when trying to build models for decision support, maybe drawing sociology and computer science together more (again: this seems to be needed in waves on twenty year centers - I met people who were there in the 50's and 60's, then lived through the AI hype of the 80's and now it seems necessary again somehow). joe Dr. Joseph Mitola III Consulting Scientist The MITRE Corporation Tampa, FL 703-314-5709 (My views are my own and not necessarily those of The MITRE Corporation nor any of its sponsors). _______________________________________________ uai mailing list uai@ENGR.ORST.EDU https://secure.engr.oregonstate.edu/mailman/listinfo/uai
