Types of Representation
This paper was written for a book in honour of Jens Rasmussen.
There may be many topics on which I do not agree with him, but this is one topic on which I respected his ideas.
As is often the case with my later papers, this is mainly a huge list of problems and complexities. Whatever may chosen as a representation of human knowledge, it must be very flexible.
Topics
1. Introduction.
2. The process operator's knowledge.
2.1. Types of knowledge.
2.2. Displaying this information to the process operator.
3. Predicate representation.
3.1. Structure and function.
3.2. Goals and working memory.
4. Pattern representations.
4.1. Uses of a pattern representation.
4.2. Diagrams of process information.
5. Using the knowledge base.
5.1. Skill and problem solving.
5.2. Bottom-up processing and override mechanism.
5.3. Limited capacity of the processes using the data base.
6. General conclusions.
References.
Types of Representation
Lisanne Bainbridge
Department of Psychology, University College London
In Goodstein, L.P., Anderson, H.B. and Olsen, S.E. (eds.) Tasks, Errors and Mental Models. A Festschrift to celebrate the 60th birthday of Professor Jens Rasmussen. Taylor and Francis Ltd., London, 1988, pp. 70-91.
1. Introduction
One of the major problems identified by Jens Rasmussen (e.g., 1979), with his usual entertaining and thought-provoking illustrations, has been the 'different forms of mental representation' a person uses in a complex task, and the relation between them. This problem immediately fascinated me, and has been the background to much of my thinking.
This question has important practical and theoretical implications. One practical aspect is that graphic displays can show patterns of structure or function which are not possible on conventional displays, and there are difficult questions about how to optimise such displays. If there are a large number of special displays for a task, then how many should be used in all? and how does one help the viewer to cross-reference between the different display formats? and how many of the displays is it necessary to be able to see at the same time? Many of the papers by Rasmussen and his colleagues have been on these topics, e.g., Goodstein (1982), Lind (1982),
There is a parallel practical problem with the many representations used by human factors engineers to model a process operator, between which, again, there are not necessarily any simple mappings. This raises questions about what we can expect from a model, and about communication between disciplines which use different modelling techniques, which I have examined recently (Bainbridge, 1986),
In this chapter I want to take a more theoretical direction, and ask about the possible form of a meta-theory of representation. If the brain can use an infinite number of representations for different purposes, what might be the basic modules which are sufficiently flexible to have this potential ? The possible answer which is presented here illustrates why understanding the process of representation is so complex.
Knowledge representation is a major topic in Artificial Intelligence. I am not going to review that literature here; I will concentrate on the representations needed for process control. In Artificial Intelligence and Expert Systems, most of the representations are concerned with language, and are language-like in form. In process control we are not concerned with the language representation problems raised by the following sentences:
The operator likes tea.
The supervisor is the operator's father-in-law.
The supervisor gives the operator a cup of tea.
The operator smiles.
A process-operation representation does not need to include everything required in a language understanding system, but does pose the extra problems of dealing with action in a dynamic world. We will use two sources of evidence in analysing these mechanisms: what must be represented, and what we know about how people use their knowledge in doing cognitive tasks. We will suggest that there are:
1. Two types of information organisation, for:
(a). The structure and function of the process.
(b). The goals and working memory of the operator.
2. Two modes of representation:
(a). Predicate (language-like).
(b). Pattern.
The discussion will be in 4 main sections:
- The process operator's knowledge.
- Predicate representation of this knowledge.
- Pattern representation.
- An operator's use of this knowledge base.
2. The process operator's knowledge
This section will make two groups of points, about
(2.1) the types of process knowledge which must be represented;
(2.2) the human information processes which simplify the problems of displaying this knowledge.
2.1. Types of knowledge
As a starting point, we will list the types of knowledge that we know the brain can deal with, because a process operator does. This is a simplified version of the list given by Jens Rasmussen (1979), in which he calls the types of knowledge 'models'. The items are listed in 5 main groups:
1. A definition of the product, and the sequence of transformations the input materials go through to make it.
2. The process limits, the constraints on its operation.
3. The process functions and behaviour.
4. The plant components.
5. The present state of the process.
2.1.1. The definition of the product : the purpose of the process
The criteria or tolerance define the product(s) made. These should be at the centre of the process operator's goal structure.
The sequence of transforms that the input materials go through transforms them to the output product. This sequence gives the next level of goals to be met in running the process. In manufacturing and chemical industries these transforms include changes in physical dimensions and state and in chemical composition, as well as conversions between mass and energy flows.
2.1.2. The process limits
The limits to the physical configuration and behaviour of the process, which are needed to operate safely and efficiently. These limits may change according to the process phase. Efficiency criteria are less important during start-up or fault management than during production. They are particularly important in batch processes, which consist of a sequence of stages. Each stage carries out a different transform, for example cooling must be completed before filtering. The end of a stage is defined as a particular process state. All these limits add to the goals to be met.
2.1.3. The process behaviour : how it transforms the materials into product
These transforms can be described independently of the physical mechanisms underlying them, as was first done in cybernetics. Three aspects are important:
(a) the general functions and the cause-effect relations between them;
(b) the process variables by which the functions are achieved, and the cause-effect relations between these;
(c) the behaviour of the cause-effect relations.
a. The functions of the process, such as cooling or filtering, and the cause-effect (cause-consequence) relations between them.
The flow of functional effects is not necessarily the same as the physical flows of material and energy in the plant. Control is also a process function, but as control is usually done either by a person or by an electronic rather than electro-mechanical device, it can be useful to analyse the control functions separately. The control functions also usually change in different process phases, such as start-up, precipitation, or fault management.
b. The process variables within each function, and the cause-effect relations between the variables.
c. The parameters of each of these cause-effect relations : gain and time constants, and the amplitude and variability of change.
These properties may also differ in different plant contexts. For example during transitions from one operating state to another, during steady-state operation, or in fault management.
2.1.4. The plant physical structure : how the functions are obtained
This can be divided into five aspects:
1. The major physical groupings of components, e.g., turbine. There is not necessarily a simple mapping between these groupings and the functions in the process.
2. The components in these major groupings.
3. The probability of failure.
4. Spatial position. The position of major structures in a workplace may be arbitrary. The position of individual components within these structures may be functional, as in a mechanical device, or arbitrary, as in an electronic circuit. The position of displays and controls on the process interface should map onto their place in the process function.
5. When a physical component can change state (for example, a valve can be either open or closed) then the state of the component is a process variable. This variable has an effect on other parts of the process, and so makes the link between structure and function (for more explanation, see Section 3.1, and Figure 1).
2.1.5. The present state of the process and the controller
The present values or states of the variables and the components : now, in the recent past, and in the near future. These may include components which are not normally part of the process function, for example nuts which are too worn and dirty to seat properly and so affect the efficiency or safety of the process.
2.2. Displaying this information for the process operator
The human information processes used in handling diagrams of process information are: (a) pattern recognition; and (b) cued memory. These make it easier to design displays, because they mean that it is not necessary to show all the process information explicitly. More centrally to this chapter, they are aspects of human information processing which must be included in our representation (of the operator).
2.2.1. Pattern recognition
People do not need a photographically accurate pattern representation. We can interpret diagrams which consist of symbols (in the sense of visual abbreviations), visual analogies of the real thing which show the important features without irrelevant details.
Patterns illustrate the Gestalt principle that 'the sum is greater than its parts'. For example, we can give a verbal description of a 'square' in the form of: four lines of equal length, joined at right angles. Give similar instructions to a graphics program and it will draw a square. A person looking at this picture of a square will be able to see properties of squares which were not in the original specification, for example that the diagonals of the square are also equal in length and at right angles. This makes it easier to design displays, because a person may read off features of a pattern representation which have not necessarily been mentioned before.
2.2.2. Cued Memory
Our memory is much more effective when we are given a cue which reminds us of associations. For example, seeing a resistor, or a symbol for an 'and' gate, evokes a memory of its function. The parameters of a differential equation imply a certain pattern of behaviour over time. Psychologists use the word 'cognition' rather than 'perception' because, when we look at a pump we see not only its physical form, we also know what it can do for us and what we can do to it.This means that in simple cases it is not necessary to display function or behaviour as well as structure, or vice versa. When developing display sets, it is necessary to think out which information should he displayed explicitly, which should he cued, and which should be left to the user's unaided memory. Of course, the availability of these memories depends on training or experience, so different display sets maybe needed to support naive and expert users.
3. Predicate representation
This chapter suggests two main modules for representing the process operator knowledge for: (a) the structure and function of the process; and (b) the goals and working memory of the operator.
2.1. Structure and function
It is easiest to start by discussing knowledge about process structure and function, although process goals are primary. We will also start with a predicate-network representation of this knowledge, as it is probably more complete, and it is easiest to represent the general structure of knowledge and the complexity of the possible interrelations in this form.
Figure 1(a). General knowledge about the category 'pump'.
Figure 1(b). Specific knowledge about a particular pump.
What is the minimum needed in a module for representing an operator's knowledge about structure and function ? Figure 1 represents some of an operator's possible knowledge about pumps, while Figure 2 below shows the basic representational module used. This is a complex node with 5 types of link :
1. Part-of links, into wholes.
2. Is-a links, into categories,
3. Attribute-with-value links, or descriptors,
4. Token-type links, from the name of an attribute to the node defining it and its attributes.
5. Cause-effect links, which may have: (a) conditionals/ discontinuities; or (b) shortcut links.
For engineers who are used to networks, it is easiest to refer to this module as a complex node with 'coloured' links. In AI it would be called a 'frame with 'slots'. In particular; it is related to object or noun frames in language understanding. We will now discuss why it has this form, and some of the additional complications of using it in representation.
Figure 2. Possible general form of a complex node representing both structure and function.
3.1.1. Structure
We will start with description of the objects and materials (and energy), in the plant.
Parts and wholes (Jens Rasmussen's 'aggregation') : part-of links.
Items are interconnected by being parts of larger wholes, by part-of links. It is not only the basic nodes that can be linked together into wholes, networks of these nodes can also be linked. In psychology these groupings may be called 'chunks'. This linking together is true not only of physical units, such as a washer - pump - turbine - cooling circuit. It is also true of any other sort of network, for example for specific memories (see Section 3.2), or our knowledge of the combination of things which are likely in particular contexts. In AI, 'frames' are part of more complex networks called 'scripts' and 'scenarios', which indicate the general types of object, event and action to be expected in general categories of a situation, such as going shopping or to a restaurant. Although these higher levels of aggregation raise important and interesting practical questions about how to display them, which Jens Rasmussen has been much concerned with, it is probably not necessary to introduce a different type of basic module for representing them in the brain.
Concepts or categories : is-a links, and attributes with values
The operator also knows about the properties or attributes of general categories of item, such as pumps. Categories are useful because they group together exemplars which have a common appearance, behaviour and function This is useful in problem solving and so on. For example, a St Bernard dog and a Scottish terrier may not have many common visual attributes, but as both bark, eat meat, and chase cats, it is efficient to group them together for purposes of predicting their behaviour.
In the representation used here, these descriptive dimensions are indicated by attribute-value links. For efficiency, information about attributes of individual members of a category, e.g., the capacity of a particular pump, should be linked to the node describing that individual. More general attributes, and their range of possible values (e.g., the possible capacities of a general type of pump), should be linked to the category. In practice, we have to learn the existence of categories with similar properties. This process of grouping is not automatic. Most of us know the 'aha' experience which comes when we realise how parts of our knowledge fit together. All the types of link in these networks have to be learned, and in both directions.
In this representation, exemplars are linked to superordinate categories by is-a links. The network of these links must be more complex than a hierarchical tree. Any item may be a member of many categories, depending on the attributes focussed on. Figure 1 concentrates on the attributes of pumps involved in moving fluids, but a particular pump is also a member of the categories 'things in the back shed', 'products made by a given manufacturer', etc. These categories are groupings of other attributes which optimise other behaviour towards the pump. Although categories are useful because of their common attributes, real categories cannot usually be defined as a concentration of attributes, so we need two types of link: is-a and attribute-value.
Figure 1 concentrates on a process operator's knowledge of pumps. A plant designer's knowledge of pumps might have other levels of category and attributes: for example, the general category of pumps might be divided into different pump mechanisms, with their general performance characteristics, and each of these into specific models, with information about their cost, availability, etc. In the designer's categories, the attributes and values are parameters which will be compared with design constraints. In the operator's case, the focus of the representation is on knowledge needed to understand and influence the process behaviour.
Process variables as attributes of objects and materials : token-type links to further information
In this representation, process variables are shown as attributes of the objects and materials. An operator may have much more extensive information about a process variable, which is used to interpret data from the process, than there is room to represent in Figure 1. For example, an operator may know:
1. The potential range of values.
2. The potential value, if any.
3. The normal pattern of values of this and related variables, if any.
4. Specific patterns of values in specific fault conditions, with 'short-cut' link to the fault identity,
5. The permitted limits to the variable value, with short-cut link to an explanation,
6. The means of altering the value (via the cause-effect links).
7. The gain, timing, amplitude and bandwidth of changes in the value.
8. The actual value in the recent past.
9. The actual value now.
10. Methods of finding the actual value now.
11. The predicted value(s), with probabilities if appropriate.
12. Methods of finding the predicted value.
(For examples of some of these points, see below in this Section.)
Because of this extensive information about an attribute, we suggest using a token-type link here. The attribute/ variable name is a 'token' for fuller information about the properties of the attribute, which are defined in another node (the 'type' node).
3.1.2. Function
Four aspects of representing function need to be mentioned: (a) cause-effect links; (b) discontinuities in the effects; (c) short-cut links from one part of the data to another; and (d) categories of function.
Cause-effect links.
The structure and function of objects have to be closely linked, though they do not have a 1:1 mapping. We have already noted the way in which a variable (as an attribute of an object) can have a causal effect on other variables. This cause-effect relation links the component into a network of causal relations which is not necessarily the same as its part-of relations in the physical structure.
Discontinuous functions.
These cause-effect links need to be able to incorporate conditions, if there are discontinuities in their effect on process state. This is important in describing both batch-processing plant operations, and fault conditions. For example: flow rate increases pressure, which is acceptable below a given limit, but above this limit causes damage. Or: cooling affects temperature, and when temperature reaches a certain level, precipitation begins.
Cause-effect chains, and shortcut links.
Cause-effect relations can be grouped together by physical linking: the effect of one is the cause of the next. These sequences are linear in simple cases, but many of the problems of understanding and controlling industrial processes arise because of branching effects, i.e., because one variable may affect, or be affected by, several others.
Shortcuts, or bypass links, appear to be an important aspect of human knowledge of cause-effect relations. For example, an operator may know that turning a particular knob affects temperature. S/he may only know it at this level, or s/he may know the causal sequence underlying and explaining this observable correlation between the process variables. Short cuts from cause to effect may underlie many important aspects of behaviour. For example, a discontinuity which causes damage may lead the designer to impose limits on all the variables prior to this in the causal sequence. The operator might work out the explanation for these limits by thinking through the causal chain to the discontinuity, or may have a shortcut link from the limit to the discontinuity. As an example : if something goes wrong with the cooling towers in a nuclear power station, the operator knows immediately that this will have an effect on the reactor, without working back through the extended causal chain by which this happens. These shortcut links occur either because they are as much detail as someone knows about a situation, or because having quick access from one part of a causal chain to another increases cognitive efficiency.
Categories of function.
Causal networks can be connected by is-a links into categories of function, for example all the ways of obtaining cooling.
3.1.3. Structure and function networks
As there are five different types of link between the nodes, the interconnections into networks can be very complex. Although these networks may not have an exact mapping with each other, they do have some mapping. These mappings provide redundancy in the knowledge. What is known in one network will support and reinforce what is connected in another network. Observation suggests that some people find it easier to remember and use the interrelations in complex knowledge networks if they 'make sense', if the structure-to-function map is not arbitrary at higher levels.
3.2. Goals and working memory
Representing the goals and working memory of the operator requires a sixth type of link: goal-method choice. This link describes the connection between goals and the choice of a method for meeting them. In this representation, working memory is a special case of a goal structure.
3.2.1. Goals
We need to discuss:
(a) why the goal-method choice link is needed; and
(b) how the goal and structure + function networks are connected.
The goal-method choice.
It might appear that goals can be dealt with simply by reversing the cause-effect links, working back from the effect required to a way of achieving it. Even if this were the case, means of meeting goals would not necessarily be obvious, as people do not automatically use knowledge in both directions. For example, we may know that boots are heavy, and that heavy things can apply pressure, but this does not necessarily mean that when we want to apply pressure we think of boots.
Reversing the cause-effect reasoning is not sufficient to account for the goal directed behaviour that people show. We need an additional type of link between goal and method. This is not because there are many ways of meeting a particular goal, and many uses of any particular method: such complexity could be handled by is-a links. Implementation of goals involves choice between methods, according to how well they meet the main and other goals. For example. how, what, and where we eat may depend not only on how hungry we are, hut also on the effect we want to have on the people we are with and on how well the requirements of a particular method, for example that a restaurant is very expensive, meet with other constraints of the present situation.
The goal networks are similar to the verb or action frames used in language representation. Figure 3 shows goals organised at several levels. These goal networks are connected in heterarchies, as there are multiple links between goals and methods (see also below).
Figure 3. A node and network for representing goals and methods
(for more information see Bainbridge 1978).
Links between goal and structure+ function networks.
The goal networks and structure+function networks may be linked in both directions. At the 'lowest' level of the goal network, methods are implemented by making actions, by pushing a button or reading a display. These actions form the link from the goal network to the structure + function network. The effects of the action may be explained by reference to the function network. In the other direction, there are links from the structure + function network to methods of finding out the value of a variable, or of changing it.
Observation suggests that people make errors when following procedures or instructions which are specified as an arbitrary sequence of actions without reasons. This is one meaning of the term 'rule-based' behaviour (see also Section 5.1). This suggests that cross-references to a goal network or a function network provide useful supporting redundancy when people are generating sequences of behaviour.
3.2.2. Working memory
Recent analyses suggest that peoples' understanding of the present state of the external world forms a type of data structure. It is related to the goal network, to what someone is trying to achieve, and it refers to long-term memory for information about the physical structure, function and potential behaviour of the process.
Figure 4. A specific example of working memory.
(for more information see Bainbridge 1975)
Analysis suggests that the 'lowest' 'sub-goals' in goal structures for information processing find:
(a) the data used in calculations or judgements; and
(b) the constraints on choice of behaviour.
Figure 4 represents part of the working memory of an expert in a specific task.We can suggest that, in a familiar task, working memory is a type of 'goal structure' which contains the information which defines the optimum behaviour choices in the near future.
The operations of working memory are more complex when people are trying to understand a new environment, as the processes involved must build a new data structure. This is part of problem solving (see Section 5.1). Johnson-Laird (1983) has discussed the processes involved in language understanding (though his terminology is confusing for human factors engineers, as he uses the term 'mental model' to designate the temporary data structure built up during understanding, rather than the background knowledge it refers to).
The data structures which are built up to represent specific circumstances can become part of long-term memory. They can be retained as memories of specific events, called 'episodic' memories by psychologists. The idea that memories are not exact replicas of previous events, but records of the processes by which we made sense of the information input at the time, was first put forward by Bartlett (1932). These episodes may become part of more general knowledge when we realise that several 'episodes' have the same properties and so form a category.
4. Pattern representation
In this section we will look at two topics, the advantages of pattern representation, and typical diagrams used for process information, categorised according to the link type represented.
4.1. Uses of a pattern representation
So far we have used a predicate (language-like) network mode for representing knowledge. But it is probably the case that most, if not all, of the information can be represented equally well, or better, in a pattern mode. There is some debate about it in psychology, but visual patterns are generally assumed to be a method of representation which parallels but is different from the network-of-predicates mode which we have used so far in this paper. Evidently, when there are both pattern and predicate representations of the same thing, there will be many cross-references between the two.
For some purposes the pattern representation may be more efficient. We have already mentioned the way in which information can be explicit in a pattern which was not in its specification. When we look at a turbine, we see configurations of parts and attributes which it would be cumbersome or impossible to describe in words. Patterns are particularly efficient at specifying the values of attributes like spatial position, proximity, shape, size, colour and intensity, and for showing interconnections. Most people find it easier to extract these attribute values from a pattern representation when necessary, rather than having to read them from a sequence of words. An example comes from elementary network problems: many people find it easier to solve simple problems by looking at a diagram of the network than by working through a mathematical description of it.
4.2. Diagrams of process information
It is evident that, with all these different types of interconnection between networks and networks of networks, the structure of knowledge is very complex. Any diagram or graphic display must extract a sub-set or slice (Jens Rasmussen's 'abstraction') of the potential knowledge, which is the primary focus for a particular task. The practical question is how to optimise the mapping between the representation and the knowledge needed in a particular task, while not taking away (the cues which help a person looking at a diagram to remember other parts of the knowledge structure) and so causing display-induced tunnel vision.
We typically use four types of visual representation for information about mechanisms, each of which describes explicitly only one of the above links. The representation conventions used in a diagram are cues from which a knowledgeable observer can infer other information. For example, 1/s is a Laplace transform symbol for integration, which is the word used to describe a certain sort of input/output relation, which is an overall term for the response of a system component, etc. When this symbol appears in familiar configurations with other symbols, this implies certain behaviour from the system of components acting together, and so on. Inversely, someone who is familiar with what is represented in an unfamiliar type of diagram may be able to infer the meaning of each of the symbols used.
Each representation gives some information explicitly, which acts as a cue to other knowledge and inferences.
An optimum representation shows explicitly:
- What the observer cannot be expected to remember, about a specific case.
- And/or the information which an observer wants to be able to manipulate easily in doing a task.
The best form for an explicit representation gives cues to what the observer can be expected to remember or infer about the general properties of the situation.
The four types of representation are:
1. Pictures of physical structure, part-of links between components.
see e.g., Figure 5 (Figures 5-10 are taken from Rasmussen, 1979, my comments)
Figure 5. The representation of physical form shows explicitly the shape, size and relative position of components, and the part-of relations between them.
An observer familiar, from everyday experience, with the behaviour and properties of these components can infer their cause-effect relations, predict their behaviour, and infer the reason for using them.
(other figures in Rasmussen (1979) which fall into this category are 1, 2, 3, 4, 6, 7, 8, 9, 12 and 13.)
2. Networks of function, cause-effect relations between variables. There are two inverse types:
- variables as nodes, transforms as links (see, e.g., Figure 6);
[Figures 6-10 all include components which need special training to understand.]
Figure 6. Signal-flow-graph of a transistor.
- transforms as nodes, and variables as links (see e.g., Figure 7.).
Figure 7. Box diagram of feedback system, using Laplace notation.
(Figures 17, 18, 19, 20, 21, 22,23 in Rasmussen (1979) are also in this category).
3. Effects of discontinuities in functions. See, e.g., Figure 8.
Figure 8. Diagram of state/event dependencies in a system.
To infer system behaviour from this, the observer needs to know what is meant by each of the different shapes.
4. Cartesian graphs showing parameters of the relations between 2 or 3 variables, attribute values of a causal link. See, e.g., Figure 9.
Figure 9. Cartesian graphs of a transistor amplifier stage.
The observer needs to know about linear and logarithmic scales, the meaning of the alpha-numeric symbols, the cause-effect relations between these variables, and what can be inferred from the shape of a curve. Visual manipulation and interpolation are involved.
Explicit descriptions combining objects and events have only so far proved possible for some simple situations, e.g., in representing the language in simple sentences, as in Figure 10.
Figure 10. Schank net representation of the objects, events and states in the sentence "John pushed the table to the wall' (Schank 1973)
(Given this information about the sentence represented, what may be the meaning of each of Schank's symbols?)
Grouping the diagrams into these categories itself illustrates how a representation can hide important aspects of the information. Anyone who has read the quoted and other papers by Jens Rasmussen will know that this grouping of his Figures ignores one of the major problems that he has been concerned with, the important practical issues in describing a process at different levels of aggregation or complexity.
5. Using the knowledge base
At this superficial and preliminary level of discussion, it is not particularly useful to say much more about the knower's use of the knowledge base than has already been mentioned in passing. But we do need to indicate the further complexities which will be needed for a full account of behaviour. These complexities are mainly concerned with the processes which use the database, rather than with the form in which the knowledge is represented.
It is important to emphasise that we have been concerned with the form of knowledge, not its content. For example, the alternative methods of diagnosing faults which have been identified by Jens Rasmussen (Rasmussen and Jensen, 1974; Rasmussen, 1981) would appear as alternative methods in the goal structure, when representing the knowledge content of a maintenance task. Similarly, the form of the knowledge base does not account for errors which arise because the content of the database is inadequate, or because there are gaps in knowledge of process structure or function, in the formulation of goals, or in the content of working memory.
We will make a few additional points about the processes which use the knowledge base in skilled behaviour and problem solving, bottom-up data input processes and limited capacity processes, and errors. These have been other major areas of concern for Jens Rasmussen, which are discussed fully elsewhere in this book.
5.1. Skill and problem solving
Jens Rasmussen's distinction between skill- and knowledge-based behaviour has had a major impact on how human factors engineers think about cognitive processes, and will be discussed by other authors in this book.
In the data structure discussed in this chapter, the goal-method choice link is the location for the flexibility with which people change between skilled and problem-solving behaviour, according to their experience and the predictability of the environment. Problem solving is the default option, when no available methods provide an adequate way of meeting the goal (e.g., Rowe, 1983), In modern psychological terminology skill may be called 'automatic' processing, and problem solving called 'controlled' processing. In this chapter, both start from goals-methods, and refer to structure + function knowledge. The important distinction is that skill uses the knowledge base networks directly, while problem solving and planning processes consider and add to them: they access them in order to build a new data structure, which is a solution or method of finding a solution.
5.1.1. Skill ['skill’ used with the general meaning of the abilities of a person who knows more than someone who knows nothing about a task, rather than to label a specific type of cognitive processing].
A skilled person knows:
1. Efficient methods for meeting the goal.
2. Optimum 'slots' for working memory in a particular task.
3. Efficient categories: groupings of items and actions with common properties relevant to the task.
4. Short-cut links around the knowledge base.
5. Appropriate attribute-values: knowing, for example, what will be the result of any method, so that trial and error, whether mentally or actually, is not necessary before choosing between methods.
Efficient knowledge makes it possible to choose 'automatically' the behaviour appropriate to the context. Skilled people can go straight through the knowledge base (the goal-method networks with associated reference information in the structure + function networks), from intention to action, without any hold-ups due to insufficient information or inadequate methods. They know all the links and expected effects. Their behaviour appears to be like an if-then production rule. (These points apply to both perceptual-motor and cognitive skill.)
Many of the current accounts of mental load are inadequate because they assume a single level of behaviour organisation, or at most a two-level organisation in which the lower level is automated and the upper level is controlled. In fact behaviour organisation probably involves multiple levels, with each level independently organised and communicating with others via the goals-method choice link. At any of these levels, behaviour may be skilled or may require conscious thought, and behaviour at levels above or below may be independently either automatic or controlled.
5.1.2. Problem solving and planning
Skilled behaviour contrasts with the behaviour of a person in an unfamiliar environment, in which it is necessary to work out what to do, and to try out alternative methods to find which works best. In doing this a person refers to and adds to both the goal-method networks and to the structure+ function networks.
Planning, mental simulation and explaining may be done using the structure-and function networks. Planning and prediction work forward through the cause-effect links and their parameters. Explaining works backwards from symptom to event.
We have already suggested that problem solving is part of the goal structure. In simple cases, problem solving consists of searching for methods which may not have been used for this purpose before, but which have appropriate outputs and requirements that can be met. Presumably (again just to mention this superficially in passing) analogical reasoning is done by a related process of comparing attributes, identifying those which are useful for the purpose, and ignoring others. An analogy may be used as a key to thinking about goals and means, attributes, or relations and structures: that is any link, node or network of nodes can be used.
This is a very incomplete account of problem solving, because it does not deal with what happens in complex cases, with the initial stages of problem definition and interpretation, or with how a discrepancy between intention and result leads to modification of behaviour or to redefinition of the problem.
5.2. Bottom-up processing and override mechanisms
We have concentrated on top-down processing, how knowledge can determine behaviour (though, as just mentioned, we have not accounted for the most complex and intelligent of these processes). Probably, there are different processes to deal with the ways in which input data can activate knowledge and override present thinking, for example, the orienting or startle response, by which we react to unexpected strong stimuli, or the not entirely different, but more subtle, way in which the state of the environment starts off thoughts about something that we were not thinking about before we noticed it. The latter can be divided into:
1. Serendipity, the way in which we do something because we notice we are in the right place, rather than because we had previously planned to (Beishon, 1974).
2. When we notice something that explains an event that happened previously, but which we did not have time to think about when it occurred.
3. When we notice that something is abnormal. (Unfortunately this is not guaranteed as a fault detection method.)
This override mechanism is an interesting one, and evidently requires a very large parallel processing ability.
5.3. Limited capacity of the processes using the data base
The limited capacity of the processes which use the knowledge base is well known. It may be that we can only think of one thing at a time when problem solving because all the mental processors are kept available in case they are needed. Limited capacity for working memory and decision making are particularly important as sources of errors. These activities require large amounts of processing capacity for building new data structures, especially when the relations between the parts of the structure are arbitrary or nonexistent, as in memory tests, or when trying to keep different plans or hypotheses in mind when making decisions. We use strategies for reducing our load of information about probability and cost, which can lead to decisions which are not optimal as mathematically defined (Sage, 1981). These error processes are another major area of Jens Rasmussen's interests which will be discussed properly by several authors in this book.
6. General conclusions
What is the practical use of such a meta-theory?
The major outcome of this chapter may be to illustrate why the representation problem (whether for, or of, the process operator), which Jens Rasmussen has always been concerned with, is such a complex one.
A meta-theory may lead to an understanding of the limits to simple formulations of the types of human cognitive processing, and may suggest why traditional behavioural empirical methods, which investigate one or two variables at controlled levels, are comparatively powerless tools for studying cognitive processes.
References
Bainbridge, L., 1975, The representation of working storage and its use in the organisation of behaviour. In Measurement of Human Resources, edited by W.T. Singleton and P. Spurgeon (London: TayIor and Francis), pp. 165-183.
Bainbridge, L., 1978, Forgotten alternatives in skill and workload. Ergonomics, 21,169-185.
Bainbridge, L., 1986, What should a good model of the NPP operator contain? In Proceedings of the international Topical Meeting on Advances in Human Factors in Nuclear Power Systems, American Nuclear Society, Knoxville, Tennessee, April 21-24.
Bartlett, F., 1932, Remembering : An Experimental and Social Study (Cambridge: Cambridge University Press).
Beishon, RJ., 1974, An analysis and simulation of an operator's behaviour in controlling continuous baking ovens. In The Human Operator in Process Control, edited by E. Edwards and F.P. Lees (London: Taylor and Francis), pp.79-90.
Goodstein, L.P., 1982, An integrated display set for process operators. In Proceedings of the IFAC/ IFIP/ IFORS/ IEA Conference on Analysis, Design and Evaluation of Man æMachine Systems, Baden-Baden, FR Germany, 27-29 September.
Jonhson-Laird, P.N., 1983, Mental Models (Cambridge : Cambridge University Press).
Lind, M., 1982, Multi-Flow Modelling for Process Plant Diagnosis and Control, Riso Report Riso-M-2357 (Roskilde, Denmark: Riso National Laboratory).
Rasmussen, J., 1979, On the Structure of Knowledge - a Morphology of Mental Models in a Man-Machine System Context, Riso Report Riso-M-2192 (Roskilde. Denmark : Riso National Laboratory).
Rasmussen, J., 1981, Models of mental strategies in process plant diagnosis. In Human Detection and Diagnosis of System Failures, edited by J. Rasmussen and W.B. Rouse (New York: Plenum Press).
Rasmussen, J. and Jensen, Aa., 1974, Mental procedures in real life tasks: a case study of electronic trouble shooting. Ergonomics, 17, 293-307.
Rouse, W.B., 1983, Models of human problem solving: detection, diagnosis, and compensation for system failure, Automatica, 19, 413-425.
Sage, A.P., 1981, Behavioural and organisational considerations in the design of information systems and processes for planning and decision support. IEEE Transactions on Systems, Man, and Cybernetics, SM-11, 640-678.
Schank, R.S., 1975, Conceptual Information Processing (Amsterdam: North-Holland).
Access to other papers via Home page
©1998, 2022 Lisanne Bainbridge
Comments
Post a Comment