The idea of using a computer program to simulate a human expert ( i.e. an informant) in anthropological analysis has been received by anthropologists with some interest, but with more caution [(.John Davis .)]. This caution is justified because to most anthropologists the inner workings of expert systems are not known; they are black boxes. But anthropologists should be interested in a model that claims to represent and use human knowledge productively, if only to evaluate that model. This paper describes some of the basic assumptions in contemporary expert systems, discusses their usefulness to anthropology, and concludes that existing expert systems are of limited interest to anthropologists engaged in qualitative research, although the general model underlying expert systems could be used productively.
Artificial Intelligence (AI) is a multi-disciplinary area in which the goal is to represent (usually human) intelligence in the modelling environment of a computer. People have worked on AI since there there have been computers. Some believed in the 'fifties that `just a few more years' would bring about a revolution in AI, but those few years have receded annually1. In the past decade there have been developments in AI that are considered by AI researchers (and others) to be partial successes. Among these is the expert system . Expert systems are computer-based models that simulate human expertise in a specific area (domain), such as a subset of medicine [(.MYCIN.)], exploratory geology PROSPECTOR, or education [(.NEOMYCIN.)]. Expert systems are claimed by AI researchers to be an important advance, and some claim implications about models of human representation of knowledge, and mechanisms of inference. [(.Barr II.)].
Qualitative analysis can be defined as identifying qualitative structures, identifying the states of those qualitative structures, and the pattern of changes (transformations) in those states2. Quantitative methods can sometimes be used to aid this process, but usually qualitative methods are exclusively used for the analysis of qualitative data and structures for which quantities proper are difficult to define.
Thom (1975) argues that all quantitative analysis assumes a firm qualitative foundation. Before they measure, people must agree that there is something to be measured, and that is a qualitative judgement. Similarly, people must agree that the measure (metric) they use is appropriate, and applicable to other phenomena3.
As an example consider per capita income. It is apparently easy enough to agree on the structure, but the metric is another issue. If currency is used as a metric, a poor family in the United States would be a wealthy one in Pakistan. The metric can be further adjusted by considering cost of living, but an acceptable level of living in the United States is not equivalent to one in Pakistan. The problem is not difficult to understand qualitatively: there are different standards in the two places. The two countries' per capita income can be compared quantitatively, but the interpretation of the comparison is qualitative. The quantitative analysis is more difficult to reconcile, and indeed is undecidable without reference to qualitative structures in the two societies.
In most cases quantitative analysis depends on continuity. To quantify a phenomenon meaningfully it is usually necessary to assume that the relation between phenomena and metric can be described by a continuous function4, since a primary goal of quantification is to provide a basis for comparison. For phenomena where the analytic focus is on states this is often misleading or impossible. In most social phenomena there is no continuous function that can adequately describe the important qualitative relationships. As an example consider income and education. These are variables that are often given a quantitative definition in social research. They are relatively easy to define, and people generally measure income in currency, education in years. But they often assume linearity and usually there will be a good correlation between them. But it will not be a perfect correlation, as one unit change in the independent variable will not result in some regular linear unit change in the dependent variable. Now this is not terribly shocking, since people do not expect that all the variation in one variable is to be explained by the other, but there is benefit in understanding the relationship between the variables by breaking the relationship into stages, and examining the conditions for moving from one stage to the next. For instance, in the USA 11 years of education is minimally better than 10 years, but 12 years is far better than 11. This is because of the local structure of American education: 11 years is pre-graduation, and 12 years is post-graduation. The graduating student has a qualitatively changed educational status, the pre-graduating student has not significantly changed status. This type of analysis helps to give a better account of interactions.
Another reason quantitative analysis must depend on qualitative analysis is illustrated in fig 1. The graph shows hypothetical data and two solutions which fit it. Solution 1 is the better qualitative fit, as the relative shape appears to be the same as the data, but is not as good a fit quantitatively as Solution 2. Solution 2 fits well quantitatively, but probably describes a different underlying mechanism altogether.
An Expert System is designed to simulate one aspect of a human expert: the ability to classify phenomena from a set of attributes. The expert system is a classification engine. It is a system that takes information about a particular case or instance within the domain of the system and produces a qualitative result (or goal state). It usually has incomplete information, and makes qualitative judgements. The basis is algorithms in a computer program plus relationships established by a human expert. This will be interesting to anthropologists if three conditions are met: the computer should arrive at the same conclusions as a native expert; it should arrive at the same conclusions as an anthropologist; and it should do useful jobs.
Expert systems, as a class of computer programs, are currently designed to reflect a general model current within the artificial intelligence community: an expert system is not simply a simulation of human expertise, but must be implemented (on a computer) in a particular fashion; it is a product of an AI culture. Ideally an expert system has two primary components [see fig. 2].
Any Knowledge Base is a set of rules describing relations between elements in the domain of knowledge. In the simplest form: [ condition(s) --> outcome ]5 The rules for deriving an outcome from a set of conditions are always formulated externally by an expert usually aided by a knowledge engineer , i.e. a specialist in transforming the expert's information into statements suitable for a knowledge base. The knowledge engineer stands to the expert as anthropologists do to their informants. The knowledge that is selected for inclusion in the knowledge base can have a variety of forms, depending on the form of the inference engine.
Most expert system designers consider it important that the rules can be easily inserted, modified, or deleted from the knowledge base, in any order. They are usually consider the rules to be weakly connected: there is no sequencing information about the order in which they can apply, and the only connections between them are the use of common terms of reference. Thus if one rule determines as an outcome that a person's residence is patrilocal, and another rule can use that residence information to draw further conclusions, the rules are connected.
An Inference Engine is a method of using the rules in the knowledge base to derive a conclusion. Using the diagram above, this might take the form:
if condition then add outcome to context
Where outcome is the conclusion if condition is true, and context is an area where knowledge is recorded if conditions are true. An outcome is often part of another condition that matches another rule. In other words, the inference engine takes the rules provided by the knowledge base, and uses internal rules of inference to draw a conclusion. The claim is that the internal rules are general to all inference. So the inference engine is a set of rules which are applied to the rules in the knowledge base.
The inference mechanism is thus critical to the outcome: it is responsible for any interrelation of elements beyond the rules in the knowledge base. It is usually based on some variant of logic, such as first-order logic, fuzzy logic [(.Zadeh 1975.)], modal logic Zeman 1973, or intuitionistic logic [(.Martin-Lo\*:f 1982.)], and also usually employs some statistical mechanisms for measurement and classification.
The inference engine is intended to be based on a general model for using knowledge and should not have special knowledge about a particular domain.
This model is claimed to be unlike the usual computer program/model structure, because the specifics are separated from the methods. This separation is made for at least two reasons6:
In most existing expert systems the knowledge base and inference engine are not terribly complex in design. The knowledge base determines the set of possible outcomes that the system can consider, and the rules for arriving at those outcomes. Although this requires great effort by the human expert and the knowledge engineer the form of representation is simple.
In many systems both outcomes and rules have an objective or subjective probability associated with with them, again derived from the human expert. The knowledge base consists of high-level structures derived from the formidable pattern matching and inference skills of humans.
An inference engine has three parts: an identification mechanism, an evaluation mechanism, and a goal mechanism. The first two constitute the inference mechanism proper; and the third is for finding efficient paths to an outcome: it does not strictly affect the outcome (unless it is poorly designed), but it selects the best condition to request data on rather than requesting all possible conditions in the knowledge base. So the goal mechanism is a search pattern through the possible conditions that apply to a case, and it is the goal mechanism that gives the expert system the appearance of performing like a human expert by requesting a minimum of information. The inference mechanism gives the expert system the judgement to announce a result consistent with the knowledge base. Most of the successful (externally validated) expert systems use some form of probabilistic model (often Bayesian) as the basis of the inference mechanism, using the probabilities associated with the knowledge base. One common goal mechanism works by finding the goal that is most likely to be true at the current time, and then finding the condition that will give the most information about that goal (as defined by the evaluation mechanism).
Consider the factors that influence the marriages arranged by urban Punjabis of Lahore [(.Fischer nd.)]. Marriages are arranged in the Punjab by the parents and other relatives of the potential groom or bride. The following factors (not necessarily in this order) appear to be the most important in the evaluation of a possible spouse:
1 zat (sometimes glossed as caste) 2 jihez (dowry) 3 intellect 4 education 5 haq (bride deposit) 6 beauty 7 izzit (honour, respect, responsibility) 8 bradarie (clan) 9 rishtidar (relative) 10 distance (from natal home)These are not Punjabi selection criteria, but an anthropologist's measurement or probe of the semantic domain of selection derived from what Punjabis say. In addition, the selection is influenced by the size of social networks and the availability especially of females, who are supposed to be invisible before (and after) marriage except to relatives.
The relationships between these measurements are complex, and they are evaluated relatively. For instance, if the zat of two candidates is different , then what constitutes enough izzit will be different in each case. In other words the state of enough izzit varies, depending on at least one other value. Amounts measured are not evaluable without other context: there is a high degree of relativity. Moreover it is probable that different people have different selection models, and one person may have more than one.
To construct an expert system based on this situation:
where the value will have already been weighted by the human operator
in terms of too little , too much , enough ,
and
where appropriate yes , no , same , and different .
The weighting is assumed to always be from the son-giving side.
This would give us a knowledge base like the following:
factor marriage 1 marriage 2 marriage 3 zat same different same jihez enough too low too high intelligence enough too high too low relative yes no yes education too low enough too high haq too low too high enough beauty enough enough too little izzit enough too high too low bradarie same different same location too far ok ok suitability yes no yes
In consultation the expert system takes in the knowledge base, creates internal rules, and answers the request, which would be for the suitability of a possible marriage [fig 3]. To derive an answer it asks the user to give values for some of the factors until it is possible to determine the qualitative result, and then makes a pronouncement, yes or no . Note it can not ask the suitability question itself, as this is the purpose of the system, but requires this for the knowledge base input to form the rules.
This example is fairly easy to follow, but has little depth because of the immense amount of analysis that is needed to set it up; for it to work it must be told to seek the correct information (the selected criteria), and that is known only after analysis. It also fails to take into account any higher-level ethnographic or ethnological knowledge; it is a purely descriptive model with no explanatory power. Additionally, the particular method described is heavily committed to a particular model in the formulation of rules, and assumes that the results are linearly differentiable; that is, that each state has a unique coordinate range in the multi-dimensional space.
Most expert systems incorporate higher-level knowledge, in the form of explicit rules in the knowledge base. The previous example can be greatly improved in performance by adding rules of the following type to the knowledge base:
1) if zat is same then izzit is enough .
2) if bradarie is same then izzit is enough .
3) if relative is yes then zat is same .
4) if relative is yes then bradarie is same .
5) if distance is too far and relative is yes then distance is ok
and so on. These kinds of rules add information about factors that cannot be taken into account in a regular statistical method. One might ask why the entire system could not be based on rules like these, freeing the system from empirical data altogether. The answer is that one can, and most working systems do. However, although the rules appear to be `higher-level', they are empirical with respect to the expert system, and provide no explanation for the outcome that is not in the rules to begin with. This defect is usually overcome in expert systems by the expert and the knowledge engineer adding comments to each rule so that when a user inquires about the reason a particular conclusion has been reached, comments are displayed for each rule invoked on the path to a successful derivation of the conclusion.
Even in this simplified model much of the complexity of the computing component is in the goal mechanism, which ideally has no analytic effect on the final outcome, whereas the inference mechanism is relatively simple, using models that are more or less in common usage in descriptive analysis. In spite of this many expert systems do often succeed in making judgements consistent with those of the human experts they are based on [(.Michie Bayes.)]. They achieve this by representing knowledge as a set of local models, made up of one or more rules, that are only weakly (and informally) interrelated8, rather than by having a single large formal model of the expert's knowledge.
Most current expert systems also have a probabilistic component. The knowledge base is for the most part entered in the form of `higher-level' rules, but objective and subjective probabilities are attached to the conditions and outcomes by the human expert. That is one way to allow the derivation of the outcome to be partial; the outcome need not be absolutely defined with respect to the knowledge base, only defined to some arbitrary degree of probability. This greatly amplifies the capacity of an expert system to classify, since it is not restricted to finding exact matches to what has been encountered before, but rather comparing as prototypes, simulating the capacity of human experts to make judgements on new cases.
There are several reasons for the success of current expert systems. First, since the local models9 as presented to the expert system are only descriptive models, and the overall system is a performance model, no internal explanation need be generated: the expert system is judged only on its descriptive performance. Second, modern statistical methods are quite powerful descriptively, so one could expect them to be reliable descriptors when used. Third, the knowledge base is created, selected and pruned by humans and consists of human expert judgements. This is also true of information supplied to the expert system while it is operating. So it is assumed that the human can answer the questions asked by the expert system appropriately and correctly. So in many ways the success of contemporary expert systems is a sleight of hand: all the human interaction in the process is taken for granted. But it is fair to say that all the expert system designers claim is to represent the knowledge of a human expert, not to create a human-like expert.
From an anthropologist's point of view the rule-based model is preferable to the statistical one, but makes no difference to the goal of the expert system, which is simply to descriptively mimic an expert. No current system can do more; expert system writers might claim psychological reality (many do not), but that is a far cry from establishing psychological reality, as the debates [see: {(.Buchler Selby , Burling Tyler.)}] over the new anthropology of the 'sixties demonstrate.
Anthropologists may still find possible significance in the general model underlying the expert system. A model of some major segment of human action need not be a single large formal model, but a series of weakly interacting local models. If these can be stated consistently, anthropologists can explore at least descriptively how the models interact with each other.
The goal of an expert system is to make qualitative judgements, to predict the state of a system relative to contextual data. However it may not be clear how an expert system can help in qualitative analysis. After all, if you have to provide the model, what is the expert system doing for you? This is not a fair argument as it applies to any computer-based aid. They do nothing you could not do, given pencil and paper and ten or twenty years. The computer in this role amplifies what can be done.
There are two more serious objections that can be raised. One is the hidden model objection which rests on what happens in the black box of the inference engine to the model or data that was entered. This is a problem only if there is no control over the identification and evaluation mechanisms in the system. In general the other mechanisms are not terribly important. For example it is not important from an analytic point of view whether the goal mechanism has a forward or backward chaining strategy. That is a description of how the information is ordered and accessed internally, rather than how it is evaluated. However, it is critical to control, or at least understand, the internal evaluation method, for the analyst is locked into the limited range of possible models that a given system can accommodate. This is strictly an issue of access to programming skill.
The second objection is to the formal or theoretical basis of the general model of an expert system. As outlined above, all current expert systems work more or less upon one general macro-method: given a list of symptoms and a list of outcomes the systems evaluates the most likely state(s) (outcome) for the system to take at each point of the analysis. The generalised expert system model attempts to achieve this global scope without explicitly laying down all the paths, rather piecing together a unique solution for each unique situation, using only a series of small, local models and a general inference mechanism as the basis. It does this not by incorporating a single exhaustive model relating all possible states to each other, but uses individual instances of information and relates them according to a weak interaction internal model. There is formal support for the weak interaction model in mathematics from {(.Thom.)}, and in anthropology and simulation from {(.Zackary.)}.
The problem with using expert systems in anthropological analysis is created by the split between knowledge base and inference engine; in general the non-programmer anthropologist can only control the knowledge base. Regardless of the type of models that the anthropologist sets up in the knowledge base, the inference model must be known to evaluate the interaction of the models as anticipated. This makes the system suspect for analysis unless one knows the inference model in detail, and is satisfied that it represents the assumptions that must be made realistically. This objection is not to the general approach, but to the fixation to a particular global model, the evaluation mechanism. The problem is not unique to expert systems, but arises in any use of simulation to test models: the result of a model always has to be tested against another model before it can be interpreted. The properties of the evaluation model must be known and consistent with its purpose. If the problem of control can be overcome then the general expert system model has potential as a means of exploring the interactions of a large number of local models towards a set of global responses, and as a method of qualitative simulation.