Intelligent Decision Support in Process Environments

Although a normative approach is necessary to provide sound foundations, utility theory has often been criticized as being too normative. This paper intends to discuss various ways of relaxing the normative framework of subjective probability and utility theory, in order to account for recent theories of uncertainty which encompass the probabilistic point of view as a particular case. This attempt aims at providing new pieces of information in the debate between the normative point of view of decision theory and the (usually) descriptive point of view of deviant uncertainty theories, often used as a basis for approximate reasoning techniques in expert systems.

In this paper I shall discuss the problem of complexity of decision rules in the framework of machine complexity. In this way, we can actually have a conceptually well-structured discussion of what one might mean by “decision complexity” and we can study the bearing the issue of complexity might have on the foundational problems of Decision Theory. I shall take a decision rule to be a Finite State Machine (FSM), and I shall define an ideal Bayesian Decision Machine (BDM). I would like to stress that I am thinking of a “machine” as a mathematical object. I shall show how BDM will meet with complexity troubles and I shall argue that it is possible to deal with them by defining decision machines endowed with “information processors” which differ from each other according to different organizations of Long Term Memory (LTM). My main claim will be that “information processors” are procedures or algorithms which denote abstract objects, namely functions, and that, whatever the underlying architecture of LTM is, it must satisfy certain very general and formal conditions to guarantee the existence of “information processing” functions on it.

Decision making in presence of weak information raises the problem of selecting the best decision model. The choice ought to be made according to pragmatic criteria. Specifically, the process of risk assessment entails various decision steps where fuzzy algebra with related generalized measures of uncertainty and preference ordering may be helpful. Two simple examples are given to show how these decision models can find applications.

The lecture deals with the problem of modelling time in decision processes. It is divided into three parts. The first part gives a survey of the concept of time in the history of thought and three concepts are identified: physical, perceived, structure of possibility. Then a short survey of the consideration given to time in ancient and modern logic. It is shown how modern model logic has produced a number of formalized languages which take into account temporal evolution and is providing a substantial back-up to the simulative languages of A.I. The second part analyses the time structure of decision making. With reference to a model “decision maker — system —ambient”, the times of the three terms of the model are discussed. The time of d.m. is shown to present all three aspects identified in the history of thought. The time of the systems or time horizon of the decision, is discussed under the aspects of perceived versus objective time, values in time, advantages and disadvantages of long-term vs. short-term decisions. The time of the ambient is considered in relation to the temporalized complexity of the world. The third part deals with the connections of time with uncertainty and with rationality. Uncertainty is discussed from the logical and informational points of view. A link is suggested between time and the fuzzy presentation of future events. Strategies for splitting the time horizon to better model uncertainties are also examined. Then the connection between the consideration given to time and the type of rationality adopted, sequential or synoptic, are discussed.

It is common knowledge that decision making is problematic. One of the clearest indications is the proliferation of decision aids, be they consultants, analyses, or computerized support systems (Humphreys, Svenson & Vari, 1983; Stokey & Zeckhauser, 1978; Wheeler & Janis, 1980; vonWinterfeldt & Edwards, in press). Equally clear, but perhaps more subtle evidence is the variety of devices used by people to avoid analytic decision making. These include procrastination, endless pursuit of better information, reliance on habit or tradition, and even the deferral to aids when there is no particular reason to think that they can do better. A common symptom of this reluctance to make decisions is the attempt to convert decision making, which reduces to a gamble surrounded by uncertainty regarding what one will get and how one will like it, to problem solving or pattern matching (Norman, in press), which holds out the hope of finding the one right solution.

Experts are often asked to assess two different kinds of probabilities. One is THE PROBABILITY THAT SOMETHING WILL HAPPEN: rain, hitting an oil well, dying in an operation, tube fracture, a total melt-down. We will call these probabilities “predictions”. Experts’ predictions are widely used in all sorts of personal and public decision making. Two examples of formal usage of expert opinion are risk analyses and expert systems used for diagnostic tasks. The second kind of probability assessed by experts is THE PROBABILITY THAT THEIR ANSWERS ARE CORRECT. We will call these probabilities “confidence ratings”. A little later we will demonstrate that predictions and confidence ratings have often been confused in the literature.

The elicitation of knowledge from an expert is practised in several fields and is of interest to various types of specialists: psychologists, pedagogues, cognitivists, engineers. There may be different goals for this elicitation and different uses made of the obtained knowledge. The psychologist, for example, expects to gain a better understanding of cognitive functioning. The elicitation of knowledge may also have practical objectives in terms of cognitive task engineering or intelligent teaching aids, which are not, however, independent of the first goal. The modalities of elicitation may vary according to the goal, but we shall be more concerned here with the general problems involved in this elicitation. We shall often refer to the elaboration of expert systems without, however, neglecting the information provided by pedagogically oriented elicitation. We shall attempt to identify the sources of difficulty in elicitation and then to determine the interest and limits of the methods involved in this elicitation. Rather than developing operational procedures we shall try to specify the difficulties inherent in these methods and examine the most favourable lines of approach. We have voluntarily avoided a direct discussion of the concept of knowledge itself.

In this chapter I combine several different paths of thought. First, I briefly review developments in the study of human information processing, in particular, those based upon parallel activation structures which operate by a form of generalized pattern matching. Second, I review some issues of human error and demonstrate that the problems are not so much in the making of the error as in the difficulty of discovering the errors, once made. Finally, I conclude with an analysis of the implications for Intelligent Decision Support Systems. The purpose of this chapter is to point out potential issues, problems, and directions of approach.

The process operator performs many tasks in his daily duty. When studying process control it is methodologically motivated to try to crystallize the essential demand that characterizes this task. Many of the researchers consider the diagnoses of the process status and the decision on the relevant operative measures the major difficulty in this type of work. This combined activity, judgement, could be defined as the critical activity, the mastery of which essentially determines the quality of the whole work.

Information processing technologies are spreading rapidly throughout the industrialized world. Can we identify cognitive prerequisites — knowledge and skills — which are particular to understanding, using, and constructing such devices? If so, we will be better prepared to deal with present difficulties of novices and workers, to educate people for future mastering of the technology, and to foresee psychological and social consequences that the dissemination of these abilities may engender. The assumption behind the talk of “computer literacy” is that such cognitive prerequisites exist and can be taught, though nobody seems to agree what they are. I shall begin by presenting and discussing a proposal about the cognitive contents of computer literacy, put forward at a conference in Houston a few years ago by B.A. Sheil (1981), a psychologist at Xerox.

Advances in artificial intelligence (AI) are providing powerful new computational tools that greatly expand the potential to support cognitive activities in complex work environments (e.g., monitoring, planning, fault management, problem solving). The application of these tools, however, creates new challenges about how to “couple” human intelligence and machine power in a single integrated system that maximizes joint performance. This paper examines some of the important issues about the use of tools to support cognitive tasks, such as what is useful advice and what is an effective combination of multiple decision makers, that are raised by the capability to produce powerful, intelligent artificial systems.

The present rapid development of advanced information technology and its use for support of operators of complex technical systems are changing the content of task analysis towards the analysis of mental activities in decision making. Automation removes the humans from routine tasks, and operators are left with disturbance control and critical diagnostic tasks, for which computers are suitable for support, if it is possible to match the computer strategies and interface formats dynamically to the requirements of the current task by means of an analysis of the cognitive task.Such a cognitive task analysis will not aim at a description of the information processes suited for particular control situations. It will rather aim at an analysis i.n order to identify the requirements to be considered along various dimensions of the decision tasks, in order to give the user — i.e. a decision maker — the freedom to adapt his performance to system requirements in a way which matches his process resources and subjective preferences. To serve this purpose, a number of analyses at various levels are needed to relate the control requirements of the system to the information processes required and to the processing resources offered by computers and humans. The paper discusses the cognitive task analysis in terms of the following domains: The problem domain, which is a representation of the functional properties of the system giving a consistent framework for identification of the control requirements of the system; the decision sequences required for typical situations; the mental strategies and heuristics which are effective and acceptable for the different decision functions; and the cognitive control mechanisms used, depending upon the level of skill which can/will be applied. Finally, the end-users’ criteria for choice of mental stategies in the actual situation are considered, and the need for development of criteria for judging the ultimate user acceptance of computer support is discussed.

When developing knowledge based systems for realistic domains the designer is faced with a complex task of planning the control structure of the system. At present most systems are developed bottom-up leading to programs which are difficult to understand and maintain. It would be desirable to be able to develop KBS’s top-down from specifications of systems functions before considering the implementation in software.In the present paper we will consider the use of a model of human decision making developed by J. Rasmussen at Riso National Labs as a tool for specification of the functions of KBS’s and their organization in several levels. We will apply the decision model in two ways. The first deal with the decisions to be performed by the KBS in terms of the domain requirements. The other application of the decision model describe the decisions to be made by the KBS in terms of manipulation of the knowledge in the KBS knowledgebase. The relevance of this application appears from the observation that the basic control cycle of the interpreter in a production system has some striking similarities with the decision model.

For the purpose of this paper an Intelligent Decision Support System (IDSS) is defined as a computer based system designed to collect, organise, process and present information to support decision making in dynamic process environments. A general system evaluation should include: (1) the quality of the system’s decisions and advice, (2) the correctness of the reasoning techniques, (3) the quality of the human-computer interaction, (4) the system’s efficiency, and (5) its cost-effectivness. I will consider the situation where there is a need to analyse the performance of an IDSS to determine the quality of its functioning, corresponding to (1) and (2) above. To do so will require a description not only of how the IDSS functions but also of how humans make decisions, since it is the functioning of man and machine together that is in focus. It also requires a characterisation of the critical aspects of IDSSs — particularly of what sets them apart from conventional (i.e. non-intelligent) systems.

If one asks oneself how to assist the human operator faced with an accident, one is led to ask two questions:1)Has he the necessary knowledge to understand the situation ?2)Taking into account time pressure, has he the opportunity to use it ? I shall attempt, based on the scientific literature and taking into account the current technological development of plants, to give some answers to these questions, in order to develop more effective technical assistance.

This paper has two related aims. First, to summarise what is known about the forms and origins of the more predictable varieties of human error. Second, to use this cognitive analysis as a basis for suggesting ways in which Intelligent Decision Support Systems (IDSS) might be employed in the process plant environment. Although the error data are drawn mainly from nuclear power plant (NPP) operations, it is assumed that their implications will generalize to other high-risk process environments.

In the design of a human machine interface for process control the designer must couple three elements as closely as possible, namely the properties of the environment, the properties of the plant, and the properties of the operator. Information and control pass backwards and forwards between these three components, and within each component information and control pass between subcomponents.

Supervisory control can be distinguished from manual control which is a closed loop, skill based action, by four important aspects:The control of highly complex, multivariable processes with mostly large to very large time constants.The discrete action patterns based on decision making processes.The variability in tasks, such as process tuning, start and stop procedures and fault management.The often vague information on the ultimate supervisory control perspectives. Hence, supervisory control tasks are different from those in manual control. Taking into account the three major process control modes which may occur in supervisory control -Normal operation, start and stop, and abnormal operation-it is of interest to classify the different tasks with reference to these control modes, as well as with reference to the three levels of control behavior as introduced by Rasmussen: Skill-based, Rule-based and Knowledge-based behavior. At the Skill-based level only manual control activities play a role; at the Rule-based level activities like process tuning and to a certain extent fault management may occur, whereas at the Knowledge-based level intelligent, cognitive activities such as optimisation, planning and fault management are thought to be.In understanding at which level of human control the different tasks can be placed best, it is instructive to study where successful human operator models are reported in literature. At the Skill-based level certainly successful control models describing manual control behavior have been reported. At the Rule-based level, less but still to a certain extent, some control and artificial intelligence models are known. At the Knowledge-based level, tasks like fault management can only be modelled, if and only if, these tasks are that well defined that they are far away from real world situations; hence only very few models have been published. So, the conclusion can be drawn that only in those cases where tasks are well defined, successful human operator models are developed. This important statement elucidates exactly the basic problem:The tasks which yield a correct description of human supervisory behavior are to be find at the Skill- and Rule-based level, they are all well defined; those which do not lead to any description of human behavior are Knowledge-based level tasks, and they are not well defined. In particular, the last class of tasks is dealing with those tasks where one needs the operator for his creativity, knowledge and intelligence.As a concluding remark the following can be said. The human operator should mainly be the adaptive, creative and intelligent supervisor who does things which can not easily or at all expected. So, those tasks which can be carefully defined, thus where modelling of the human behavior is expected to be successful, can easily be taken away from the operator. Often intelligent MMIf can help the operator to support him to do the job. Whether this is wise or not is dependent on factors such as the need for training, for learning, for building up an Internal Representation of the process, etc. So the final and major question becomes: to what extent should one automate, or, to what extent should one build in artificial intelligence in man-machine system interfaces?

The study of Human-Machine Interaction is approached here with the aim of discussing the needs of an analyst for the correct design of a complex system. The two terms which we have introduced above merit some further definitions: system is in our terminology the set of all those features which are needed for the performance during the time of a certain function. These features include machines, humans, interfaces, computers, procedures, etc.;complexity is here referred to the many dependencies which intentionally or not are built within the various items of the system.

The main reason why humans are involved in controlling and supervising process environments of all kinds is the fact that humans are and have to be ultimately responsible for decision making in critical situations. This has to be considered in the design of technological systems and of their organizational environment. The pure automation approach fails, at least in complex systems, because the interface to the human can only inappropriately be added afterwards. The allocations of tasks and responsibilities between different technological, human, and organizational system components strongly influences the behaviour of the overall system under normal and emergency situations. This is particularly important with respect to the decision making capabilities when time pressure, uncertainties or risk have to be mastered.

An information system, in our sense, arises when a person, with some goals and intentions, recognizes that her/his state of knowledge is inadequate for attaining the goal. To resolve this anomaly, the person goes to some information provision mechanism (IPM). The IPM, in general, consists of a knowledge resource and some intermediary access mechanism. Thus, the information system consists of a user, who instigates the system, a knowledge resource (KR), which contains the texts which might be relevant, and an intermediary mechanism (IM), which mediates between the user and the KR. Some typical systems of this type are in student advisory services, social security benefits offices and bibliographic retrieval systems. The system as described is certainly a joint cognitive system (Woods, this vol.; Hollnagel, this vol.), and the IPM has the basic characteristics of a classic decision support system.

I have two concerns in this paper. The first is to raise the point that some decision support displays might have a syntax of their own which obscures the very data they are supposed to enhance, just as a misleading analogy may lead one to overlook an important aspect of a problem- In other words, enhanced displays may sometimes sow the seeds of their own destruction. The second concern is to suggest that supposed expertise is sometimes heavily dependent upon certain display formats- We may be in danger of making overly general claims for the depth of operators’ knowledge, when the truth lies ultimately at a far simpler, more perceptual level. The answers may lie in a detailed examination of the problem itself and its context, making an ‘ecological’ approach to decision support necessary (Gibson, 1969; Neisser, 1976).

Machinery condition monitoring and fault diagnosis using a fault matrix and process deviation approach is described. In particular the application of computer aided diagnosis using an expert system based on artificial intelligence is considered. A prototype fault finding system has been developed and tested. Some examples from this system and some more recent results from an ongoing research programme are outlined.

Interpretation and classification of sensor data in many applications can still not be automated and must be performed by human operators. Representative examples are sonardata as well as infrared or X-rayed pictures. The task of the operator in such applications is on one hand the extraction of features from the picture and on the other hand the interpretation of these features using an internal knowledge base.

This essay is a summary of the views that were expressed at the NATO Advanced Study Institute on Intelligent Decision Aids in Process Environments during a panel discussion. The panelists were D. Dubois, Université Paul Sabatier, Toulouse; B. Fischhoff, Decision Research, Eugene; P. Garbolino, Scuola Normale Superiore, Pisa; G. Volta, CEC-Joint Research Center, Ispra; and W. Wagenaar, University of Leiden, Leiden. The panel was chaired by the author, who also assumes full responsibility for the contents of this summary.

It is important to have a realistic appreciation of what intelligent decision aids can and cannot be expected to accomplish in process industries. Accordingly, the panel discussion concentrates mainly on decision making by plant operators in the various circumstances that may occur. In this respect, four inter-related issues can be distinguished and dealt with in a sequence. First, is the need of producing some kind of knowledge representation for decision making in realistic environments. Second, is the meaning of descriptive and prescriptive models for decision making, in order to assign a proper role to intelligent aids. Third, is the assessment of values that different intelligent decision aids may have in improving the operation of engineered plants and systems. Fourth, is the identification of cases and situations where intelligent decision-support systems (IDSS) would find their best use.

Throughout the papers and discussions expressions like consultant, prosthetic, tool, amplifier, replacement, advisor, assistant have been mentioned. But what can we say about what constitutes decision support, much less ‘intelligent’ decision support? What knowledge is resident in a decision support system? What is the relationship between the machine and human elements of the ‘human IDS ensemble’?

Different perspectives lead one to see different aspects of the same thing. The papers and discussions at the Advanced Study Institute (ASI) reflect this truism. I intend to draw out some of the varving perspectives on intelligent decision support systems (IDSS) and to examine some of the themes and issues that arise naturally out of these perspectives. I could ask which of the perspectives is the most productive, but that question would likely produce rather unproductive arguments. Instead I shall attempt to profit from taking various perspectives and thereby gaining a more complete view of the problem at hand.

As both the written papers and the edited versions of the panel discussions show, a considerable number of issues were raised during this ASI. It is difficult to summarise these on a few pages if one wants to do more than just enumerate them. And to do them all justice could easily produce a small book of its own. The solution chosen here is to discuss three of the main issues in some detail. The choice reflects the opinion of the three ASI directors and should not be seen as the opinion of the ASI participants as a whole. We are, however, convinced that these issues would rank among the most important on anybody’s list even though they might not be the top three ones.