EIAxpert: An Expert System for screening-level   EIA




Fedra, K., Winkelbauer, L. and Pantulu. V.R. (1991)
Expert Systems for Environmental Screening.
  An Application in the Lower Mekong Basin.   RR-91-19. International Institute for Applied Systems Analysis. A-236l Laxenburg, Austria. 169p.



3   Expert Systems for EIA

Expert systems, an emerging technology in information processing and decision support, are becoming increasingly useful tools in numerous applications areas. Expert systems are man--machine systems that perform problem-solving tasks in a specific domain. They use rules, heuristics, and techniques such as first-order logic or semantic networks, to represent knowledge, together with inference mechanisms, in order to derive or deduce conclusions from stored and user-supplied information.

Application- and problem-oriented systems, rather than methodology-oriented ones, are more often than not hybrid or embedded, where elements of artificial intelligence (AI) technology, and expert systems technology in particular, are combined with the more classical techniques of information processing as well as the approaches used in operations research and systems analysis. Here, traditional numerical data processing is supplemented by symbolic elements, rules and heuristics, in the various forms of knowledge representation.

There are numerous applications where the addition of a quite small amount of ``knowledge'' in the above sense, for example, to an existing simulation model, may considerably extend its power and usefulness and at the same time make it much easier to use. Expert systems are not necessarily purely knowledge driven, relying on huge knowledge bases of thousands of rules. Applications containing only small knowledge bases, of at best a few dozen to a hundred rules, can dramatically extend the scope of standard computer applications in terms of application domains, as well as in terms of an extended non-technical user community.

Clearly, a model that ``knows'' about the limits of its applicability, what kind of input data it needs, how to estimate its parameters from easily available information, how to format its inputs, run it, and interpret its output will require not only less computer expertise from its user, it will also assist the user with domain expertise in the application area.

Artificial Intelligence and expert systems

In discussing a domain as loosely defined as expert systems, it may be useful to present a few definitions selected from the literature, to set the stage and introduce the jargon. Equally instructive are the essentially graphic definitions that are available (Figures 3.1 and 3.2).

Expert systems, or Knowledge Based Systems, are a loosely defined class of computer software within the more general area of AI, that go beyond the traditional procedural, algorithmic, numerical, and mathematical representations or models, in that they contain largely empirical knowledge, for example, in the form of rules or heuristics, and inference mechanisms for utilizing this form of information to derive results by logical operations. They are fashioned along the lines of how an expert would go about solving a problem, and are designed to provide expert advice. Like any other model, they are sometimes extreme simplifications and caricatures of the real thing, i.e., the human expert.

However, definitions or functional descriptions of expert systems and claims to the expert system category of software cover a broad spectrum, ranging from fairly modest to rather optimistic parallels to human, or even super-human, performance:

Most existing expert systems work in analytic domains, where problem solving consists of identifying the correct solution from a pre-specified finite list of potential answers ... (Merry, 1985)

Expert systems are computer programs that apply artificial intelligence to narrow and clearly defined problems. They are named for their essential characteristic: they provide advice in problem solving based on the knowledge of experts (Ortolano and Perman, 1987).

An expert system is a computer system that encapsulates specialist knowledge about a particular domain of expertise and is capable of making intelligent decisions within that domain. (Forsyth, 1984).

An expert system ``handles real-world complex problems requiring an expert's interpretation [and] solves these problems using a computer model of expert human reasoning, reaching the same conclusions that the human expert would reach if faced with a comparable problem.'' (Weiss and Kulikowski, 1984).

There are, however, even more demanding definitions. In their description of MYCIN, one of the classic expert systems, Buchanan and Shortliffe argue that an expert system

...is an AI program designed (a) to provide expert-level solutions to complex problems, (b) to be understandable, and (c) to be flexible enough to accommodate new knowledge easily.'' (Buchanan and Shortliffe, 1984).

One of the more extensive definitions and more optimistic descriptions comes from Hayes-Roth:

``An expert system is a knowledge-intensive program that solves problems that normally require human expertise. It performs many secondary functions as an expert does, such as asking relevant questions and explaining its reasoning. Some characteristics common to expert systems include the following:
  • They can solve very difficult problems as well as or better than human experts

  • They reason heuristically, using what experts consider to be effective rules of thumb and they interact with humans in appropriate ways, including via natural language

  • They manipulate and reason about symbolic descriptions

  • They can function with data which contains errors, using uncertain judgemental rules

  • They can contemplate multiple, competing hypotheses simultaneously

  • They can explain why they are asking a question

  • They can justify their conclusions.'' (Hayes-Roth, 1984).

Obviously then, there seems to be no generally accepted definition of what exactly is an expert system. Descriptions and definition in the literature range from rather narrow automata selecting pre-defined answers to better-than-human reasoning performance in complex problem domains. There is, however, general agreement that an expert system has to combine:

  • A knowledge base, that is a collection of domain-specific information;

  • An inference machine, which implements strategies to utilize the knowledge base and derive new conclusions from it (e.g., modus ponens, forward chaining, backward chaining);

  • A knowledge acquisition mechanism that elicits information required not only from the user, but also from domain experts so as to initialize the knowledge base,

  • An explanation component, that can, on request, explain the system's inference procedures,

  • and a conversational user interface that controls and guides the man--machine dialogue.

Obviously, an expert system must perform at a level comparable to that of a human expert in a non-trivial problem domain.

In summary, a concise description of AI would be the art or science of making computers smart and expert systems could be described as smart problem-solving software.

Basic concepts behind expert systems

What makes expert systems different from ordinary models and computer programs? Rather than trying to define differences in any formal way, it may help to introduce and discuss some of the basic concepts and approaches used in expert systems.

Expert systems are alternatively referred to as knowledge-based systems. Knowledge representation, therefore, is one of the fundamental concepts and building blocks in expert systems.

Knowledge is represented in various forms and formats, following different paradigms. The more commonly used forms include rules, attribute--value lists, frames or schemata, and semantic networks. A brief but comprehensive introduction to knowledge representation is given in Chapter 3 of Barr and Feigenbaum, 1981.

Formal logic and propositional calculus offer a basic form of knowledge representation. Well-defined syntax and semantics and expressive power make it an attractive option.

A proposition, a statement about an object, is either TRUE or FALSE. Connectives permit the combination of simple propositions. The most commonly used connectives are:

Rules of inference, such as {\em modus ponens}, allow the derivation of new statements from given ones: if X and X --> Y are TRUE, then Y is also TRUE:

The rules of propositional calculus, extended by predicates, allowing more complex statements with more than one argument, quantifiers such as for all $(\forall)$ and there exists $(\exists)$, and inference rules for quantifiers, result in predicate calculus (Barr and Feigenbaum, 1981). Adding the idea of operators or functions leads to first-order predicate logic, and this, restricted to so-called Horn clauses corresponds to the syntax of Prolog (Clocksin and Mellish, 1984; Bratko, 1986).

Probably the most widely used format, and also the most directly understandable form of knowledge representation are rules, also referred to as productions or production rules, or situation--action pairs. They are close to natural language in their structure, and they are familiar to programmers used to classical procedural languages such as FORTRAN or C: IF ... THEN ... ELSE is easy enough to understand. Examples of rules would be:

RULE 1010320 #encroachment corridor by forest type
IF     landuse              == forest
AND    forest_value         == high
AND    [ vegetation         == rain_forest
       OR vegetation        == dense_forest ]
AND    wildlife             == abundant
THEN   encroachment_corridor = very_large
ENDRULE

RULE 1010532 #USLE soil_erodibility
IF   [  soil_type ==  very_fine_sandy_loam
     OR soil_type == silt_loam ]
AND  soil_organic_content < high
THEN soil_erodibility = high
ENDRULE

Obviously, the terms used in rules can be more or less cryptic and require proper definition and interpretation in the system:

RULE 1010201 #degradation by watershed class
             #and land requirements
IF        project_country     == Thailand
AND       [ watershed_class   == WSC1
          OR watershed_class  == WSC2 ]
THEN      Impact  = major
ENDRULE

Structured objects are another popular means of representation of information knowledge. They are known as Schemas (Bartlett, 1932); Frames (Minsky, 1975); Prototypes or Units (Bobrow and Winograd, 1977); or Objects in many languages or language extensions, e.g., SMALLTALK (Kay and Goldberg, 1977); LOOPS (Bobrow and Stefik, 1983); or FLAVORS (Moon and Weinreb, 1980).

Frames allow combinations of generic and specific information, where the former can be inherited within a hierarchy of frames, consisting of classes, super- and sub-classes, and instances. As a data structure, frames for example can combine declarative and procedural components. Slots as units of descriptions can hold attribute--value pairs, but also function specifications and of course reference to other frames.

Another form of representation is by means of {\em semantic networks}, which consist of nodes representing objects, concepts, and events, and links or arcs between the nodes, representing their interrelationships (Quillian, 1968). A well-known example of an expert system using semantic networks is PROSPECTOR, dealing with mineral prospecting (Duda, Gashnig and Hart, 1979).

A specific and very important feature of expert systems is the inference engine, i.e., the part of the program that arrives at conclusions or new facts, given the primary knowledge base and information supplied by the user. The basic principle was already hinted at above in the introduction of predicate calculus.

There are two basic strategies, namely forward and backward chaining. Forward chaining implies reasoning from data to hypothesis, while backward chaining attempts to find the data to prove, or disprove, a hypothesis (Forsyth, 1984). Since both strategies have advantages as well as disadvantages, many systems use a mixture of both, e.g., the Rule Value approach (Naylor, 1983).

For many practical purposes, developers use expert systems shells and special development environments rather than basic languages such as C, C++, LISP, PROLOG, or SMALLTALK. While shells may offer the advantage of easy use and ready-made structures and formats, they sometimes tend to restrict the user to specific forms of representations, and, for the more complex and comprehensive ones, are expensive. For a more recent survey and discussion of selected software for expert systems development see Ortolano and Perman, 1987.

Expert systems in environmental modeling

here is a rather extensive and very rapidly growing literature on AI and expert systems, starting from the, by now almost classic, four-volume Handbook of Artificial Intelligence (Barr and Feigenbaum, 1981; Barr and Feigenbaum, 1982; Cohen and Feigenbaum, 1982; Barr and Feigenbaum, 1990). Recent review articles concentrating on environmental systems and engineering, and water resources in particular, are for example, Ortolano and Steineman, 1987; Rossman and Siller, 1987; Hushon, 1987; Gray and Stokoe, 1988; Beck, 1990.

The number of expert systems being described in the literature are many. The number of operational systems, in everyday use for practical purposes, however, seems to be rather small, in particular when looking at an area such as environmental impact assessment.

Of the 29 systems compiled in the table on page 26, almost all are in the R \& D stage; little or no information exists on successful practical applications on a routine basis. This, however, does not make expert systems different from the vast majority of simulation and optimization models developed in the field.

Another feature is that a large number of systems have been developed for operational applications rather than planning, in particular in the wastewater treatment area. Groundwater systems, especially those related to hazardous waste management problems, are another obvious focal point. Finally, there are several Intelligent Front-End systems, i.e., model selection or parameter estimation tools.

APPLICATION DOMAIN Contact or reference
Initial screening and scoping of envl. impacts US Army Electronic Proving Ground, Ft. Huatchuca, Arizona
Screening of environmental projects ESSA Ltd., Vancouver
Environmental resource evaluation Portugese Ministry of Environment
Consultative system for environmental screening ESSA Ltd., Vancouver
Wetland management U.S. Fish and Wildlife Service, Ft. Collins, Colorado
Environmental technical info. system University of Illinois
Environmental assessment system Institute for Environmental Studies, Free University of Amsterdam
Multiple-use watershed management (MUMS) Hushon, 1987
Groundwater flow analysis Andrew Frank, Dept. of Civil Engg. Univ. of Maine
Groundwater contamination (DEMOTOX) Ludvigsen, Sims and Grenney, 1986
Groundwater vulnerability (AQUISYS) Hushon, 1987
Well data analysis (ELAS) Weiss, 1982
Water resources laboratory aide Bob Carlson, Dept. of Civil Engg; Univ. of Alaska
Oil spill simulation Antunes, Seixas, Camara et al., 1987
HSPF simulation advisor (HYDRO) Gaschnig, Reboh and Reiter, 1981
Mixing zone analysis (CORMIX1) Doneker and Jirka, 1988
Input parameter estimation for QUAL2E Barnwell, Brown and Marek, 1986
Hydrologic model calibration J.W. Delleur, School of Civil Engg; Purdue Univ.
Parameter estimation for runoff model (EXSRM) Engman, Rango and Martinec, 1986
Advisor for flood estimation (FLOOD ADVISOR) Fayegh and Russell, 1986
Model selection for surface water acidification Lam, Fraser and Bobba, 1987
Trickling filter plants (sludge Cadet) Catherine Perman, Dept. of Civil Engg; Stanford University.
Anaerobic digester Michael Barnett, Dept. of Envir. Science and Eng; Rice Univ.
French water treatment plant Pierre Lannuzel CERGRENE/ENPC
New York water treatment plant Steve Nix, Dept. of Civil Engg; Syracuse University
Activated sludge plants Deborah Helstrom, Dept. of Civil and Environmental Engineering, Utah State University
Activated sludge diagnosis Johnston, 1985
Water system loss Steve Maloney, CERL
Sewer system design Lindberg and Nielsen, 1986

Compiled from Ortolano and Steineman, 1987; Rossman and Siller, 1987; Hushon, 1987; Beck, 1990; Gray and Stokoe, 1988.

Types of applications

There are several types of expert systems applications in any particular domain: they range from purely knowledge-driven systems or ES proper, to ES components in an intelligent front-end, to fully embedded or hybrid systems. Each of these systems have their specific characteristics, use, and problems. As with any attempt at classification, real things do not neatly fit into square boxes, but it helps to structure the discussion and appears to satisfy a basic need of the scientific mind.

An expert system proper would be a purely rule-based system, relying on a sizable knowledge base. As such, it is based on a largely empirical ``model'' or a qualitative, causal understanding of how things work. In the world of water resources modeling, that would put it in a class with the universal soil loss equation rather than a finite element model based on an albeit simplified version of the Navier--Stokes equations. What it describes or models is not ``the system'', but an expert's understanding of the system, in particular, his problem-solving approach and strategies.

There are only a few purely knowledge based systems that do not contain a substantial conventional component. Some of the operation and control systems, in particular in the wastewater treatment area, seem to fit into this category. Further, a large number of systems are being developed for hazardous waste site assessment and related topics, such as permitting or waste site management, e.g., WA/WPM Generator (Paquette, Woodson and Bissex, 1986); RPI Site Assessment (Law, Zimmie and Chapman, 1986); GEOTOX (Mikroudis, Fang and Wilson, 1986; Wilson, Mikroudis and Fang, 1986); DEMOTOX (Ludvigsen, Sims and Grenney, 1986); or SEPIC (Hadden and Hadden, 1985). Reviews of these systems can be found in Ortolano and Steineman, 1987; Rossman and Siller, 1987; Hushon, 1987.

An intelligent front-end is a user-friendly interface to a software package, which enables the user to interact with the computer using his or her own terminology rather than that demanded by the package.'' (Bundy, 1984).
What they can do, among other things, is to avoid or minimize misuse of complex models by less experienced users.

The QUAL2E Advisor, FLOOD ADVISOR, HYDRO, CORMIX1, or EXSRM are all examples of this type of application. Systems of this nature help a user to select the appropriate model to be used, assist in specifying input parameter values, and provide interpretation of the model's output (Rossman and Siller, 1987).

The QUAL2E Advisor (Barnwell, Brown and Marek, 1986) is a rule-based system, built with a commercial expert system shell, M.1. The system suggests appropriate parameter or input values for coefficients used in modeling stream temperature, the type of hydraulic model used and its associated coefficients, and biological oxygen demand removal, sediment oxygen demand, and reaeration rate coefficients. Appropriate values are suggested in a question-and-answer session, where information about stream characteristics that can be easily obtained, e.g., by visual inspection, such as shape of channel cross-section, slope and depth, nature of stream bed, bank vegetation, are used to classify the river and estimate corresponding coefficients.

Hybrid systems, finally, represent an integration of classical algorithmic techniques with AI and expert systems methods. The basic idea of an expert system is to incorporate into a software system expertise, i.e., data, knowledge and heuristics, that are relevant to a given problem area. However, classical simulation models are a rather powerful class of ``heuristics'' (after all, most of them incorporate a considerable amount of expertise, and they are empirical to a more or less obvious degree, even if they claim to be ``physically based'').

Models could also be viewed as a special case of production rules. In any case, they are useful in many situations, and are even more useful if combined and extended with rule-based components that add a considerable amount of flexibility in problem representation as well as estimation and evaluation methods.

Much of the above also holds true for the intelligent front-end system, and any attempt at a clean-cut classification will be found wanting; hybrid systems with embedded AI components would simply have several, in fact many, ``micro expert systems'' integrated into the overall software package. They rely on a number of disjunctive and specialized knowledge bases in different representation formats, depending on the domain and its most natural form of representation.

Several examples of integrated hybrid systems that also contain water resources models are described e.g., in Fedra, Weigkricht and Winkelbauer, 1987; Fedra, Li, Wang et al., 1987; Fedra, 1986; Fedra, 1988. The basic philosophy and early examples are described in Fedra and Loucks, 1985; Loucks, Kindler and Fedra, 1985; Loucks and Fedra, 1987.

Benefits from expert systems

AI and expert systems technology are certainly an intriguing new development in computer science that hold great promise for better applications. However, like any other method, they do not offer universal solutions and need a thorough understanding of their requirements and limitations for proper use.

By and large, expert systems are empirical systems, based on a more or less explicit, and usually qualitative, understanding of how things work. A perfect example of an ideal application area is law, or in the context of water resources, water rights and allocation problems. In water resources modeling, however, there is a substantial amount of physically based modeling, where an understanding of how things work can be expressed quantitatively. Much of our quantitative ``understanding'' is still empirical and not based on laws of nature (Darcy's law is an empirical formulation but then, physicists would argue, so is Schrödinger's equation).

However, it is important to realize that expert systems are certainly no substitute for many time-tested methods and models, but should be seen as complementary techniques which can improve many of these models. Obvious applications related to numerical models are in data pre-processing, parameter estimation, the control of the user interface, and the interpretation of results. There are certainly enough arts and crafts components in numerical modeling that open attractive opportunities for AI techniques.

While there is certainly some application potential for a purely knowledge-driven system in classifications and diagnosis tasks, the most promising area of application is in coupled, embedded, or hybrid systems, such as intelligent front-ends, intelligent interfaces, and modeling support rather than new models themselves. When integrated with data base management and interactive color graphics, AI concepts can help to shape a new generation of powerful but truly user-friendly ``smart'' software that actually gets used in planning and management.

AI applications are no longer restricted to expensive special-purpose hardware, but are increasingly supported on standard workstations and powerful PCs. With this wide accessibility, and an increasing number of affordable software tools, we may well be at the beginning of an exciting era of new developments and applications.


© Copyright 1995-2016 by:   ESS   Environmental Software and Services GmbH AUSTRIA | print page