KIVA: an archaeological interpreter
By
Jitu Patel and Arthur Stutt
HCRL Technical Report No. 35
June, 1988
Abstract:
We prent a system which simulates the interpretation of archaeological sites. Previous systems in archaeology have performed lower-level classificatory tasks. The prent system - KIVA - combines data-driven and expectation-bad reasoning to produce an overall interpretation of a given site (i.e., a cultural profile). In this approach, a model of a typical archaeological site is utilized in order to provide the constraints necessary to deal with the inherent uncertainty of archaeological interpretation.
Keywords:Archaeological interpretation, expectation-bad reasoning, reducing uncertainty.
The modern kiva ... is an artificial cave, the ceremonial centre of the
village, in which there is also a small hole in the ground, symbolic of the
place of emergence. (Encyclopedia of Religion, Macmillan, 1987)
1. Introduction
The u of expert systems in archaeology is relatively new. Given the lack of certainty which attends interpretations of the past, this is hardly surprising. On the other hand, it is this very uncertainty which, becau it lends itlf to heuristic reasoning, makes this domain of particular interest to the expert system rearcher. The few systems that do exist are mostly confined to simple classificatory tasks (e.g. Bishop & Thomas, 1984; Ennals & Brough, 1982). Even though there has been relatively little progress in developing such systems, there has been considerable discussion of their merits and demerits (Baker, 1986). One of our aims is to show that expert system technology has something to offer to archaeology (and indeed, vice versa), by prenting an archaeological expert system which is capable of performing some of the higher level interpretive tasks engaged in by archaeologists.
This paper prents a system which emulates the reasoning process of archaeologists in interpreting archaeological sites. The model of interpretative reasoning ud is derived from the work
of J-C Gardin (Gardin 1980, 1987) in which he puts forward what he calls a logicist analysis of archaeological reasoning. This approach takes archaeological reasoning to be a process of applying transformations to initial propositions (Po) to arrive at terminal propositions (Pn) or interpretations via a ries of intermediary propositions (Pi). A system bad on this approach should be able to take a description of a site or group of sites and produce a t of one or more interpretations or cultural profiles, as output. A cultural profile gives a complete account of the various activities entered into by the inhabitants of the site in terms of chronology, technology, ecology, economy, social organization and ideology.
In brief, according to this model, archaeological interpretation has the following components (e figure 1):红色男爵
1a.Classification of features. The term "features" encompass the various aspects of archaeological sites both man-made and natural (e.g., pits, ditches, walls etc.).
1b.Classification of finds. "Finds" are mobile features which are either man-made or natural such as bone fragments.
2.The reconstruction of past human activities in terms of activity areas and
their associated activity. An "activity area" is a significant area of a site at
which identifiable human activities (e.g., cooking or hide-working) were carried out.
3.Cultural interpretation. That is, the creation of an interpretation or cultural
profile for the site as a whole which includes a determination of the technology, subsistence, social organization and religious or other beliefs of the occupants of the site.
While this model of archaeological reasoning is adequate for some purpos such as the reconstruction of the reasoning of archaeologists as exhibited in their written texts (Gardin 1987), its esntially bottom-up nature fails to capture an important component of archaeological reasoning; that archaeologists go onto a site with a certain expectation of what they will find. This expectation is ud to interpret both individual artifacts and their relationships to other finds. Indeed, if an unexpected find is made some other expert may need to be called in. This aspect of reasoning is not confined to archaeological interpretation. For instance, our expectations of what will be the ca guide our everyday inferences.
水克火
Thus, the type of reasoning we are dealing with has similarities to the process of abduction and deduction which is common in, for example, fault diagnosis (e.g.
Jophson, Chandrakaran, Smith and Tanner 1987). However the expectations in archaeological reasoning are derived from a large body of common-n knowledge rather than a tightly constrained body of, say, electro-magnetic theory.
The importance of the omission of an expectation-bad component is shown in the following example.南达科他级战列舰
1.1 An example of archaeological interpretation
优秀老师
Suppo that in a given site we have found the following:
早安心语正能量• a ring of stones
• a pit within the ring of stones
• a spatially related stone artifact with sharp elongated edge
建筑平面设计During the process of classification we can u the finds to infer, for example, that there is a firepit on the site (a pit surrounded by a ring of stones) with a knife (a stone artifact with a sharp elongated edge) related to it. Using this information and our knowledge that knives may be ud for cooking or
hide-working we can provide a reconstruction (or Pi propositions to u Gardin's term) of the area around the fire-pit as either a hide-working area or a cooking area. This, in turn, can provide the basis for a cultural reconstruction at a higher level, for example, that the economy of the site is predominantly hunter-gatherer.
1.2 Discussion of the example
We can e from this example that the system has two alternative explanations of what the activity area was ud for (cooking or hide-working). How is the system to proceed? In the next ction we review some other approaches to dealing with uncertainty which have been applied in other domains. The subquent ction outlines our own solution for the archaeological domain.
后鼻韵母怎么读1.2.1 Other approaches
The standard means of resolving the sorts of problems include various conflict resolution methods, certainty factors and fuzzy logics (e.g. Zadeh 1988). Issues to do with the techniques are not addresd here since they are best thought of as a low level issues in the design of inference engines for knowledge bad systems. Moreover, in the ca of certainty factors and fuzzy logic, we think that the mathematical and logical techniques employed may tend to give a feeli
ng of objectivity to what is after all a process rooted in the highly subjective assignment of numerical values.
One further technique for dealing with uncertainty which is more cloly related to our own is that of Paul Cohen (1985) in which he takes an endorment bad approach to uncertainty. In Cohen's approach, endorments or explicit reasons for belief are attached to rules and facts and propagated by the system during reasoning. The endorments are ud to decide between conflicting possibilities.
1.2.2 Our approach
The above example illustrates one form of uncertainty in archaeological reasoning; i.e., that which aris becau more than one interpretation of the data will always be possible. As in everyday reasoning, the inferences that are made are usually plausible rather than certain. This is becau there is inevitably a gap between the material evidence and the interpretations placed upon it. While there have been attempts to produce law-like generalizations in archaeology (e.g., Schiffer 1976), the have mostly been confined to the lower levels of interpretation and there is some doubt about their applicability even at this level. At the level, say, of religious belief, it is unlikely that there can ev
er be a law which states that finds of type T mean that religious practice of type R was carried out at site S in the past.
西上辞母坟
However, archaeological reasoning is uncertain in another way. Since there is a limited supply of data available and this may be destroyed in the process of excavation which precedes interpretation, the evidence upon which the archaeologist bas his or her interpretation will always be radically incomplete. This incompleteness in the data will result in uncertain inferences.
We have chon to implement an expectation-bad model as a solution to the problem of uncertainty in archaeological reasoning. This approach has clo affinities to the model-bad approach advocated by Clancey (1986) as well as to the techniques for resolving uncertainty in the domain of language understanding (e.g. in Hearsay-II, Erman, Hayes-Roth, Lesr and Reddy 1980). The approach has been chon since it ems to capture the actual practice of archaeologists. When an archaeologist interprets a site s/he has a model of the kind of site which determines what s/he expects to find. Thus in the example given above, the archaeologist, thinking that s/he is excavating a Pueblo Indian site, will expect a particular range of artifacts and features. If a find or feature is wrongly identified at some early stage, the application of the fine details of the model will generally rve to correct initial misconceptions.
In archaeological reasoning, therefore, the correct interpretation is the one which subsumes as many of the finds as possible without infringing any constraints on the combination of possible activities on a site. We prefer the term expectation-bad to model-bad since in the actual implementation there is only an implicit model of an archaeological site. The components of this model are split between rules which act as constraints on what components a site can have and rules which can merge possible
interpretations. We also think of the system as reducing rather than resolving uncertainty since, in many cas, there will be no one correct interpretation of an archaeological site.
Thus, in our system, KIVA, no values, either numerical or symbolic, are propagated. KIVA builds up all possible solutions and, from its knowledge of a typical site, picks out the best solution (or solutions). In the above example the system could apply a t of constraint rules which includes the knowledge that all Pueblo sites have a cooking area. Thus it could determine that it is better to believe that the activity carried out in this particular area was cooking since the area has a firepit and no other area of the site has evidence for this necessary component of a site of this kind. Furthermore, from its knowledge that hide-working was never carried out in an area rerved for cooking, it can determine that cooking was the only activity carried out.
Unlike the alternative approaches to uncertainty, the expectation-bad approach deals with both the multiplicity of contending interpretations and the incompleteness of the data. In the ca of the former, the site model can be ud to reduce the number of interpretations. In the ca of the latter, the model can be ud to make plausible assumptions which will allow reasoning to proceed even when data is missing. In our implementation, we have concentrated on the former.
Moreover, since the knowledge which is ud to lect the possible interpretations is reprented explicitly (in what is, in effect, a distributed model of a typical archaeological site) the knowledge of how the system reached its decision about its reasoning is available for possible u in explanation (e Clancey, 1983; for an argument approach to explanation, e Stutt 1987, 1988). The very uncertainty of the interpretations possible in archaeology makes the provision of an adequate explanation capability even more important than in other domains. The explicit reprentation of the knowledge ud to reduce the uncertainty will enable future extensions of our system to include a fully adequate explanation capability.
A description of the actually implemented archaeological reasoning system is prented in the next ction. Section 3 of this paper prents a sample run of the system.
2. KIVA - the archaeological interpreter
KIVA1 is designed to interpret hypothetical archaeological sites bad on the current understanding of American Indian Pueblo cultures (Longacre 1970, Schiffer 1976). The 1In this paper, "KIVA" refers to the knowledge ba, and "kiva" to an activity area. In the rest of
the paper bold-italics are ud to indicate units in the KIVA knowledge ba
