Lisanne Bainbridge

Ergonomics/ human factors needs to know how to design to support people who are doing complex tasks, and classic ergonomics does not offer many of the concepts and techniques needed. This paper indicates that classic ergonomics is based on inadequate assumptions about the nature of the cognitive processes underlying complex behaviour. This paper makes suggestions about the cognitive models which are required, based on evidence about the knowledge used by, and behaviour of, people doing complex tasks. Knowledge is structured in many ways by such people, and they must have cognitive processes for handling these structures. People's behaviour in complex tasks is adaptable, as a function of details of the context, and this adaptability must be inherent in the mechanism which generates the sequence of behaviour. Both factors suggest that cognitive processing is done by processing modules working within a context, rather than by a strict hierarchy or a set sequence of processing stages.

The first section of this paper is about the assumptions often implicit in the use of hierarchies and cognitive modelling. The next three sections discuss various aspects of organisation in complex tasks. The first discusses 'horizontal' organization, how subordinate items could be grouped together to form a superordinate item. In hierarchical terms, this describes the relation between any two neighbouring levels in a hierarchy. The next section discusses 'vertical' organisation within one cognitive activity, that is the nature of the relations between levels throughout a multilevel hierarchy (or what may appear at first sight to be one). The last of these three sections discusses sequential organisation of different types of cognitive activity. All these discussions indicate how to change the models of mechanisms for complex behaviour which are used by ergonomists. A final section introduces some of the implications for ergonomic practices.

2. Implicit assumptions about hierarchies and models

The words hierarchy and model are often used with implicit meanings. These are made explicit here, to clarify the contrast with points made later in the paper.

The structures of knowledge used in complex tasks (see below for examples) are often described as hierarchies. Here a hierarchy means a structure in which items at one level of detail are combined together to make a superordinate item, over several levels. There is a general assumption about the meaning of this word 'hierarchy' (apart from the original religious meaning of people ranked in order or importance). Most people use the word loosely, but it is taken implicitly to mean an inverted branching tree structure. More formally, the parts may be assumed to be connected together to make items at a higher level by the Boolean operators 'and' and 'or'. It is also implicit that these operators are sufficient to account for the organisation, and that the same organisational principle applies at all levels of the tree.

The classic model for the sequence in which people carry out cognitive activities consists of a set sequence of stages - attend, perceive, choose response, execute response, and so on. There are many versions of this model, which has been prevalent in post-Broadbent (1958, figure 7) experimental psychology and in ergonomics. It is related to ideas from engineering, particularly information theory, and control theory. There are many implicit assumptions. Processing consists of a set one-directional sequence of processing, from stimulus reception to response execution. Each processing stage responds to, and only to, the output of the previous stage. The output of a stage is a simple if-then mapping of its input. Other knowledge is referred to, if at all, only later in the processing, the processing is input driven. Rasmussen's (e.g., 1980) stepladder model is a more extended model of this type. It includes stages with more complex input-output relations such as 'define task'. And it allows some of the stages to be omitted in short cuts. However it does not contain any mechanism for context effects, or any flexibility in the basic sequence of stages. The rest of this paper will illustrate the importance of these aspects in complex tasks.

3. 'Horizontal' organisation

If we ask what are the principles of organisation used in building the structures of knowledge and behaviour often described as hierarchies, we find that their nature implies, ironically, that these structures are best represented, not as hierarchies at all, but as independent modules which process different levels of detail and operate in parallel. A much richer and more interesting range of organisational principles than Boolean operators is used. This paper will outline some of them. It makes no claims to completeness. The types of structure discussed are the ones needed to account for complex tasks such as those in this special issue, in particular the knowledge which some industrial process operators have of the world they are interacting with, and how to change it (see Bainbridge 1988, 1993) (although simpler examples will usually be given).

The first part of this section is a general discussion of the ways in which items can be related together. Two subsections then discuss particular types of organisation : is-a (classification) structures, and part-whole structures.

The general types of link between knowledge items which are found in process operation are associations, descriptors, and groupings of items between which there is some sort of 'inclusion' relation, i.e., the items together make another item at a superordinate level.

Examples of associations are :
* arbitrary codes used to convey meaning, e.g, blue = steam;
* empirical knowledge used without deeper understanding, e.g., if this is displayed, then do that action.

In industrial process operation, examples of descriptors, or attribute-value pairings, are :
* probabilities and costs/payoffs of events;
* production targets or plant constraints not linked to a deeper explanation;
*metaknowledge about the general attributes of some behaviour, e.g., how long it takes, how much effort, how rewarding it is, how reliable, etc.

Winston et al. (1987) distinguish between four types of semantic relation : attribution, attachment, ownership, and inclusion. This paper focuses on inclusion relations. In an 'inclusion' relation, the way in which the items are organized together itself conveys usable information, and implies the cognitive processes needed to handle it. Winston et al distinguish three types: spatial, class, and part-whole. However, when process operators know about spatial positions, these spatial relations may convey other information. As examples: if one part of the plant is burning, parts next to it are also in danger; or items may be grouped together on the interface because they have the same function, or because they should be used in sequence. So spatial position can be an attribute defining a classification or a part-whole relation. This section now discusses these two in more detail.

3.1. Is-a (classification) structures

Is-a structures are ones in which items are grouped together into a superordinate category because they have similar properties. For example dogs and cats are both members of the category 'domestic animals' because they live in the home, provide companionship, etc. The items which are members of a category are grouped for convenience, but the category itself is a cognitive construct, and the items remain independent elements.

Items could be grouped together (for example) because they have the same general appearance, the same expected behaviour, the same function, or should be acted towards in the same way. Categorisation means that life is not an infinity of special cases. It gives processing efficiency for handling known cases, and makes it easier to deal with unfamiliar items, because, once they have been assumed to be a member of some category on the basis of some properties, then they can also be assumed to have the other properties of that category, and other members of the category can act as analogies. All this implies that there is more to categorisation than a concatenation of items which can be given the same name. There are two general points.

A category is an information structure which enables certain types of information processing. The superordinate item provides the name of the category. It also provides the important attributes of the category and their expected range of values, which indicate what one can check an item for in the present or expect from it in its past and future.

If a category is a structure of items which go together because they have common properties, then how can a category be defined? It seems that particular features of appearance or behaviour are only important or defining in relation to a particular task or context. For example, what features of 'dogginess' allow one to recognise something as a dog? As a dangerous dog? As a potential champion of its breed? It is not possible to define categories a priori, independent of a particular task or context in which the items will he thought about or used. Indeed, one aspect of cognitive skill is to learn which items form the categories which are useful in a particular task. This paper will not discuss categories of is-a structure in more detail.

3.2. Part-whole structures

In part-whole structures, the parts go together to make a whole that is more than the sum of its constituent parts. These groupings of items typically either make up a whole entity, such as the parts of a typewriter, the workforce of an industrial company, or the variable values in the present state of an industrial plant, or they meet some purpose, such as the collection of actions involved in plastering a wall.

There are three features of part-whole structures. The subordinate items are necessary to make the superordinate item. The lower level items may or may not have to be in a particular configuration to make the superordinate item. For example, a heap of head, body, and legs does not make up a dog, a particular spatial configuration is also essential. And the lower level items may or may not be independent when they are part of the higher level item. For example, one could say the tail of a dog is not the same thing when it is not connected to a dog, but a car carburettor remains one whether or not it is in a particular car. (These examples illustrate that these distinctions are not always clear or stable, but may depend on a particular way of looking at things in use at a particular moment.) This interdependence may be another important aspect of human skill. Leplat (1989) describes the way in which subunits of activity, whether perceptual or muscular, change in character when they become part of a skill, so that it may no longer be possible, or easy, to use them as separate behaviours.

There is a large number of different ways in which items can be grouped together into a larger whole. Some of the ones which appear in an industrial process operator's knowledge are suggested here. Different types are named, but the nature of each is not analysed.

* concatenation, but no necessary configuration: the parts are necessary, but not in a particular pattern;

* logical operators: when 'and' and 'or' are sufficient to describe the relations between items, as in a fault tree;

* perceptual configurations, e.g., 'figural goodness', as in the Gestalt principles for what makes items look as if they go together and make a whole;

* various types in which a specific simultaneous or spatial configuration is necessary, such as:
a. a configuration of conceptual entities, e.g., the functional structure of an industrial plant;
b. a configuration of physical objects, such as the physical structure of an industrial plant;
c. a configuration of variables, such as the cause-effect structure of an industrial plant;
d. a configuration of variable values, such as a plant state (e.g., a pressure/ temperature combination);
e. a configuration of activities, such as the movements in a multidimensional physical skill (e.g., tennis or football);

* various types in which a specific sequential or temporal configuration is necessary, such as:
a. ordering, e.g., 'eht' are letters, but 'the' is a word;
b. temporal sequences, such as the pattern in which the value of a variable changes over time.

These types of configuration have been distinguished because one could use a different descriptive tool to make an optimal representation of each. Winston et al. (1987) also distinguish: stuff-object, portion-mass, member-collection, and place-area, as part-whole relations.

To anticipate the next section somewhat, many part-whole configurations are complex, multilevel, and contain a mixture of these simple types. Several types of configuration may be relevant and interlinked, or one type of configuration may be implemented by another (e.g., functional structure - causal structure - physical structure). (There may also be is-a categories with particular properties within these general types.) Examples of more complex configurations are:

* episodic memories for specific past incidents or cases;

* expected sequences of events;

* combinations of events and actions in:
a. the life history and character of the process;
b. working procedures/ methods/ strategies;
c. scripts and scenarios.

It is not clear whether there is any necessary limit to the number of different types of configuration which might occur in cognitive processing as a whole. Also, in the same way as for is-a groupings, it seems that it is not often possible to give a task- or context-independent definition of what are the necessary elements making up a larger entity, or their necessary configuration. It is possible to give an independent rigorous definition for only some, for example logical operators, or control-theoretic descriptions of changes in variables over time. For most of the structures, the only definition for what is necessary, in the elements or the configuration, is that 'it works'. It is an interesting question for cognitive science to devise representational mechanisms which have this sort of flexibility, combined with a representation of 'wholeness'.

4. 'Vertical' or multilevel organisation

In a 'branching tree' hierarchy, the same principle of horizontal organisation is used at all levels of the tree. If the organisation is done by Boolean operators, then it is possible to deduce what is at a higher level of the tree, given complete information about what is at the lowest level. A key question is whether this is also true for the part-whole 'hierarchies' of knowledge used or referred to by cognitive processing, and if not, what this might imply.

This section will discuss the organisation of behaviour within one activity, using as examples two typical activities in which cognitive processes build up a structure of representation in working storage. In 'understanding', this structure represents the person's inferences about what underlies the input information. In 'implementing goals', working down from a main goal to a level of detail which can be implemented, the structure represents the plans and actions which have been assessed and chosen.

Both these activities have been frequently described by a hierarchy. However, the evidence suggests that the mechanism underlying each of these activities is modular, with each module processing a different level of detail, rather than strictly hierarchical. It also suggests that processing does not necessarily start, either from input information or from stored knowledge, but is a mixture of the two. This implies a contextual mechanism. More evidence on this is surveyed in the last sub-section.

4.1. Understanding input information

The examples will be taken from reading, because this is simple to describe. Psycholinguistic experiments show that letter/word/sentence do not form an additive hierarchy, as at first thought, but instead different principles of organisation operate at each level of detail. For example, someone who did a lifetime of research on the optimum visual presentation of only one digit would not find that 3, 5, 6, 8 are easy to confuse because they have features in common, so the optimum design of a single digit is not independent of the task context, the ensemble of alternatives from which it needs to be discriminated. No amount of research on perception of individual letters would lead one to predict the word superiority effect, that letters are processed differently when they are in words. One could not predict, from research on single words, that active sentences are easier to read than passive ones. Research which focused only on single sentences would not lead one to realize that their meaning can change when they are embedded in different paragraphs. (For example, the semantic markers in 'It went for a walk' as a single sentence make the reader assume that 'it' is animate. But if this sentence was in a paragraph of idiomatic English, 'it' could be something inanimate which had disappeared.) These sorts of finding suggest that within the processing at each level of detail there is a local knowledge-base of alternatives, and local principles of organisation, which do not apply at other levels. So written language processing is now not described as a hierarchy but as a set of independent modules each of which processes a different level of detail (Rumelhart 1977).

Because there are different principles of organisation at each level of detail, or within each module, it is not possible to deduce what happens at one 'level' from complete information about another, as it is in a Boolean hierarchy. Indeed, modules may be identified because they have different knowledge bases and principles of organisation, and so different definitions of completeness.

The examples above also show that processing at one level of detail is not just based on information from the next lower (i.e., greater) level of detail, but is also affected by processing at higher (i.e., less detailed) levels of organisation. This means that one level of detail is not superior or subordinate to another, but instead the modules for different levels of detail operate in parallel, and provide the context for each other's operation.

'Module' and 'context' are both words which different authors use with different meanings. In this paper a 'module' means a structure which contains its own cognitive processes, organising principles, and reference knowledge, and which can be used independently, in the sense that it can occur before, after, or simultaneously with other types of processing. Which types of processing module are used will depend on a particular task. This use of the word does not imply any particular brain localisation.

In this paper the word 'context' is used to summarise all the types of material which are available for a cognitive module to work on or with. This usually includes three main sources: information from the environment; the output from other modules currently in working storage (e.g., the current understanding and plans); and stored knowledge. In process operation, stored knowledge can be of a large range of types (Bainbridge 1988, 1993), such as production targets, task constraints (e.g., time available, safety limits), how the plant works, and how to understand and operate it, as well as personal aims and preferences. This surrounding context evokes, delimits, and defines the pertinence of potentially available knowledge and other inputs (see figures 1 and 3 in Bainbridge 1993).

4.2. Implementing goals

In the organisation of motor behaviour we find the same features as in perceptual organisation: the independence of 'levels' of organisational detail, and the adaptable direction of influence. Meeting goals has traditionally been thought of as a hierarchy of goals and sub-goals (e.g., get to work = get to station + take train + get to office). Motor activity has often been described as a hierarchy of motor programmes. But surely, when playing golf, cricket, or a musical instrument, it would he so much easier to repeat a good movement if that were true. In many sports, such as tennis or football, the actions required have to be so adaptive in detail to the context that a hierarchy of predefined motor programmes is very unlikely. And people who play the piano or violin may have had the experience of playing a well known sequence of notes with a different fingering, without conscious volition. Locally organised self-regulating modules give a more flexible mechanism than a rigid hierarchy. A rigid goal hierarchy which said 'get to work' - 'use a train' would be helpless if no train were available. But a mechanism which separates goals and means ('get to work' - 'among means of travel or getting to work, using a train is the most convenient') allows a search for other alternatives when one of the means is not available.

So again, the notion of a strict hierarchy in biological and cognitive control systems has been replaced by the notion that a definition of the required output, at a level of greater detail, is transmitted to that level, but not detailed instructions about how this goal should be met. Also both directions of processing are involved. If for some reason it is not possible to meet a sub-goal at a 'lower' level of detail (e.g., there is a train strike, in the 'get to work' example), this may mean that it is necessary to revise the sub-goals or plan at a higher level. It looks as if the process of implementing goals is also best represented by independent modules which operate in parallel on different levels of detail, rather than by a simple treelike hierarchy.

Anyone who has tried some sort of hierarchical task- or behaviour analysis may have discovered that there is no task- or context-independent definition of what should be at a given level, nor of the number of different levels needed to make a complete description. As in the previous discussions, the definition of 'completeness' at a given level of detail is that the goal is met, or the device works, not that specific methods have been used.

4.3. The direction of influence in processing

This subsection starts with a short example of processing from a more complex task, which also shows that the direction of influence in processing is not only from input to output. The section then summarizes the main aspects of inner-initiated processing in complex tasks.

As an example, Table 1 in Bainbridge (1991a or 1992) describes the operators thinking in response to an alarm at the beginning of an incident at the Prairie Island nuclear power plant, as analysed by Pew et al. (1981). Everything which the operators use in interpreting the situation, after the initial orientation to the alarm, is not displayed but is supplied from the operators' knowledge:
* alternative hypotheses about possible explanations for the displayed information;
* knowledge about what other information to look for to check their hypotheses, and how to find it;
* expectations about what will happen next;
* actions appropriate to their hypothesis about what is happening in the plant.
The operators build up in their working storage, using their knowledge base, a structure of understanding about what is happening in the plant. Analysis of a longer sequence of their activity (e.g., Table 2 in Bainbridge 1992) suggests that the operators carry forward their interpretations, predictions and plans. Studies of teamwork, e.g., Reinartz and Reinartz (1989) suggest that teams build up a communal structure of this sort of understanding.

Direction of influence in processing is also an area in which many different terms are used. Some of them are: data information (i.e., external world) driven or knowledge (internal) driven; environment or goal; bottom-up or top-down; lower, i.e., more detailed or higher, i.e., less detailed. This paper will use 'input' and 'inner', in the hope of avoiding the varied accretions of meaning around the other terms.

Inner-initiated processing has at least four important characteristics.

4.3.1. Active attention: The knowledge base is not just something which is referred to on the way from stimulus to response, as in a sequential-stages model. The knowledge base suggests alternative hypotheses about what may underlay the input information (which in complex situations is nearly always incomplete). It also suggests what other information is needed to test between these alternative hypotheses, and where to look for this information in the environment. Recent versions of the sequential stages model include attention (the selection of material to process) within a person's internal sensory buffers. In real complex tasks, people also direct their attention in the external environment for the information that they need.

4.3.2. Adding knowledge to the interpretation, which was not in the input information:
The classic small example of this comes again from reading:
'It was Harry's birthday. Mary went to the store and bought a kite'. Who was the kite for?'
Answering this question involves knowing about toys and presents, which are not mentioned in the original phrases. These are inferred from the knowledge bases related to the original 13 words; they build on and expand what is in them. Most of the types of knowledge described in Section 3 might be used in making such inferences.

4.3.3. Anticipating, and if necessary acting to change, future events: Several types of knowledge are used to predict, and work out how to improve, future events. As an example, in the Pew et a!. op. cit. example (e.g., Table 2 in Bainbridge 1992), after the operators had identified the cause of nuclear radiation (a major leak) they predicted the sequence of events which would happen as a result of the leak, and took action to minimize the anticipated bad effects.

4.3.4 Organising future behaviour : In complex tasks people do not only work out how to translate main goals ('keep the nuclear plant in a safe state') into sub-goals which can be implemented. If they have several simultaneous responsibilities (main goals) they may multi-task, meeting the multiple goals by working on each in turn a bit at a time (Reinartz, 1989). They may also organise their future behaviour by identifying the need for enabling actions (HolInagel 1993), or by working out a plan (e.g., Beishon 1969). Planning involves processing of inner origin, as it involves imaging and comparing alternative potential sequences of events, which have not actually occurred.

All these aspects imply the need to include both context effects, and adaptability in the direction of influence of processing, in any model of the cognitive processes used in complex tasks.

5. Sequences of activity

Section 4 discussed the limits to using a strictly hierarchical account to describe the cognitive processing at various levels of detail within one cognitive activity - understanding information, or making an action, for example. This section expands the types of cognitive activity to include ones which are used in trying to understand and operate a complex dynamic external world. It also looks further at models for the sequence in which people carry out different cognitive activities. Evidence about sequencing behaviour supports the same conclusions about the processes which need to be modelled as were made in Section 4, that processing modules work within the context of the output of other modules, and that sequences of processing activity by people doing complex tasks are adaptable in ways which a sequential stages model would be inadequate to produce. People do next what is most appropriate as a function of details of the context. This need for changes in the type of model used to account for complex behaviour is also discussed in Bainbridge (1991 a), Grant (1992) and Hollnagel (1992a, 1993).

The first subsection here points to some evidence on the adaptability of processing sequences. The final part suggests how a sequential stages model could be a special case of a modular contextual model.

5.1. Sequential adaptability

Evidence shows that sequences of behaviour in complex tasks are adaptive. For example, in industrial process operation the sequence of thinking done by operators consists of small sections, in each of which they think about one part of the task. These sections are modular, in the sense that they can be done before or after any other type of task thinking (see, e.g., Table 3 in Bainbridge 1992), so they appear to be independent.

Depending on the task, the modules carry out information processing of various types. Table 4 (16k) in Bainbridge (1992) suggests a list, including : identify, infer/ review present state, review/ predict events, predict state, review/ predict task goals, evaluate state, review action availability and effects, review compensatory events, choose best action, identify need for enabling action, plan future activities, make action, monitor action effects, and monitor changes. Note that these are cognitive goals, the goals of thinking, which in complex tasks are an intermediate step in meeting task goals. This sort of list is necessarily typical, rather than complete, because what it is appropriate to do will depend on the needs of a particular task. (Hopefully the ergonomists' job will be simplified by identifying classes of task and associated processing.) An important part of human skill might be to develop task specific processing modules, with their associated reference knowledge and working storage.

The evidence suggests that there are not many constraints on the order in which these modules are used. In its strongest form, a sequential stages model consists of a set number of stages carried out in a standard sequence. Evidence from complex behaviour suggests instead that the order in which types of processing are done depends on the context at a particular moment. The evidence from process operation is summarized in Bainbridge (1992) and from human reliability by Hollnagel (1993). This class of mechanisms were first suggested for process operation in Bainbridge (1973, 1975) and for reading by Rumelhart (1977).

5.2. Integrating sequential and contextual models

Despite the limitations of the sequential stages type of model as an account of complex behaviour, it must he useful to account for something or it would not have been successful for so long. It is possible to suggest that it is a special case of the contextual type of model, that sequential processing is the default behaviour of a contextual mechanism when there is no context for it to set up and operate in. Three examples will illustrate this.

5.2.1. Psychological laboratory experiments: Laboratory experiments often deliberately eliminate context effects, and test independent stimulus - response pairs. It would be ridiculous to suggest that such experiments should not be done. From the theoretical point of view, these experiments provide much food for fascination, and in practice they have provided most of the classic ergonomic design recommendations. However, this type of experiment is limited as a methodology for getting information about, and therefore driving models to account for, more complex types of cognitive processing (see Rasmussen 1993).

5.2.2. Tasks in which a context-free strategy is optimum: The topographic strategy used by the maintenance technicians studied by Rasmussen and Jensen (1974) is an example. Because the technicians work on a wide variety of different types of equipment, it is efficient (from the perspective of mental effort) for them to use a fault-finding strategy which does not depend on specific knowledge about any particular equipment. They do use working storage, as Rasmussen and Jensen describe, but only to build up a yes-no picture of the acceptability of individual components in a piece of equipment, not to understand the state of the equipment relative to its particular purpose.

5.2.3. The first response to a new situation in a complex task: When an industrial process operator first comes on shift, or when an unexpected alarm goes off, the operator does not have to start without knowledge of the specific context. However, studies of fault diagnosis, for example (Decortis 1992), show that operators do not necessarily go through the expected sequence of inferring the state of the plant in detail, i.e., fully diagnosing the fault, before choosing an action. They may start with some sort of holding action which prevents the situation from getting worse, before they take the time to work out in detail what has gone wrong.

Contextual models will only be needed to explain data in which there are contextual effects.

The type of contextual model suggested in this paper, which focuses on more complex cognitive processes such as understanding and planning, is itself incomplete. An ergonomist also needs a model of human cognitive processing which draws attention to, and accounts for, the features which increase or decrease a person's capacity for doing a task, such as memory limitations or problems with translating from one code to another, as these are central to making detailed ergonomic design recommendations. Barnard (1987) provides a modular model which does focus on these aspects.

6. Some implications for ergonomic practice

The key argument of this paper has been that evidence on the organisation of reference knowledge, and of sequences of behaviour, suggests that the underlying mechanisms of complex behaviour are modular and contextual. Cognitive processing is done within at least three aspects of context: the task relevant knowledge and goals, and the results of previous thinking, as well as the input information. These are used to build up an information structure in working storage, which represents the person's current understanding of what is happening and what they should do in relation to it. This paper has not suggested possible mechanisms for these context effects, see Bainbridge (1992) and Hollnagel (1993).

Why should this change in models be important for ergonomists ? It has implications for both research and practice. In research, apart from issues of model development, there are also major gaps in our knowledge of how people do complex tasks, and of methodologies for investigating behaviour which is mostly unobservable. The papers in this special issue illustrate some of the problems and approaches. A contextual model focuses attention on such topics (among others) as how people: infer what is going on behind a complex interface; anticipate future events; build up a simultaneous overview of a large portion of a task to provide the context for task decisions; plan future activities; understand why a complex piece of equipment did not do what they were expecting; or develop and maintain the cognitive skills used in doing these activities.

These are all factors about which we need more information in order to design complex equipment. Classic ergonomics focuses on optimising pre-specified (in particular repetitive) work. Design recommendations have been primarily concerned with minimising physical effort, and with simple aspects of mental effort, such as minimising the use of low capacity information processes and optimising the use of high capacity ones. Now, in many tasks, people are using flexible equipment to do unanticipated or unprespecified tasks, and a different approach is needed.

Although the implications of the contextual approach for ergonomic design have not been much discussed, there are comments in the literature. It is only possible here to indicate what is needed. This note points out issues in the design of display systems, task allocation, human error, and task analysis, for industrial process operation tasks.

Five issues illustrate some of the implications for display system design:

1. Because all aspects of a display are processed in parallel, the features of a display which support detection, discrimination, identification and interpretation all need to be equally effective (if needed) For example, suppose a new display to support interpretation is designed using circle and hexagon as shape codes. It may then fail in tests because the shape discrimination is difficult, rather than because the new principle of interpretation is wrong.

2. Gestalt organization principles are important in the design of graphic displays, but allowing for them is difficult to incorporate in computer-based design assessors.

3. Active attention should he more effective when information is displayed in stable locations, as then automated information-acquisition eye-movements can be learned.

4. Overview displays should help people rapidly to update or recover their understanding of what is happening (though much research is needed on how to design these well).

5. Because cognitive processes are done in context, not in isolation, this raises serious questions about the use of subtask specific displays: subtask specific displays may make it easier to do a particular subtask, but less easy to integrate this subtask into an understanding of the larger task which this is part of.

Some of these issues are discussed more fully in Bainbridge (199lb)

That human cognitive processing operates within a context also needs to be taken into account in allocating responsibility between people, automatic devices, and procedures. Here are some illustrative points:

1. People need time to build up their understanding of a complex situation before they can take effective action. This has implications for the design of manual takeover from automated systems.

2.Although more research is needed on this, the available evidence suggests that people only learn to build up easily accessible reference knowledge bases, and effective overviews of what they are doing, by doing a task themselves, not from demonstrations or lectures. This has implications for both the nature and length of training for complex tasks.

3. Procedures should state what is the aim of any action they prescribe, so the user can understand the context of what they are doing.

Some of these issues are discussed more fully in Bainbridge (1983, 1990).

The nature of contextual processing is also linked to difficulties in predicting human cognitive error rates, as discussed fully by Hollnagel (1993). Failings in the nature of the information structure built up while doing a task, or in the organisation of an effective sequence of behaviour, are important error mechanisms which are not dealt with in most error analyses, and are discussed by Bainbridge (1993).

Putting any of these design issues into practice needs to be based on a task analysis. The contextual approach suggests that the context, that people doing a task need to build up and maintain, should be a focus on task analysis. Some of the implications of this have been discussed by Grant and Mayes (1991) and by Bainbridge (1989a).

All these points illustrate that the development of our understanding of the contextual nature of complex cognitive processes is not an abstruse theoretical issue but has extensive practical consequences.

References

BAINBRIDGE, L. 1973. An analysis of a verbal protocol from a process control task. unpublished Ph.D. Thesis. Department of Psychology. University of Bristol.

BAINBRIDGE, L. 1975, The representation of working storage and its use in the organisation of behaviour. in W. T. Singleton and P. Spurgeon (eds) Measurement of Human Resources (Taylor and Francis, London), 165-183.

BAINBRIDGE,L. 1983, Ironies of automation. Automatica, 19, 775-779

BAINBRIDGE, L. 1988. Types of representation. in L. P. Goodstein. H. B. Anderson and S. E. Olsen (eds) Tasks, Errors and Mental Models (Taylor and Francis, London). 70-91.

BAINBRIDGE. L. 1989a, Cognitive Task Analysis. Department of Psychology. University College London.

BAINBRIDGE. L. 1989b. Types of hierarchy. types of model. in L. Bainbridge and S. J. Reinartz (eds) Proceedings of the Workshop on Cognitive Processes in Complex Tasks (TUV Rhineland. Wilgersdorf, Germany) December, 75-85.

BAINBRIDGE, L. 1990, Will expert systems solve the operator's problems? In R. A. Roe. M. Antalovits and E. Dienes (eds) Proceedings of the Workshop on Technological Change Process and its Impact on Work, 9-13 September, Siofok, Hungary. 197-218.

BAINBRIDGE, L. 1991a. Organising principles in hierarchies and models. in Proceeding of the Third European Conference on Cognitive Science Approaches to Process Control. School of Psychology, University of Wales at Cardiff, 26 September. 81-90.

BAINBRIDGE, L. 1991b, Multi-plexed VDT Display Systems, in G. R. S. Weir and J. L. Alty (eds) HCI and Complex Systems (Academic Press. London). 189-210.

BAINBRIDGE, L. 1992, Mental models in cognitive skill: the case of industrial process operation. in Y. Rogers, A. Rutherford and P. Bibby (eds) Models in the Mind (Academic Press, London), 119-143.

BAINBRIDGE. L. 1993, Difficulties and errors in complex dynamic tasks. Department of Psychology, University College London.

BARNARD, P. 1987, Cognitive resources and the learning of human-computer dialogues, in J.M. Carroll (ed.) Interfacing Thought: Cognitive Aspects of Human~Computer Interaction (MIT Press, Cambridge, MA).

BEISHON, R. J. 1969, An analysis and simulation of an operator's behaviour in controlling continuous baking ovens. reprinted in E. Edwards and F. P. Lees (eds), 1974. The Human Operator in Process Control (Taylor and Francis, London), 79-90.

BROADBENT, D. E. 1958. Perception and Communication (Pergamon Press, Oxford).

DECORTIS, F. 1992. Processus cognitifs de resolution d'incidents specifiés en relation avec un modele theorique, Ph.D. Thesis, Faculty of Psychology. University of Liege, Belgium.

GRANT, A. S. 1992, A context model needed for complex tasks, in Proceedings of the Second Interdisciplinary Workshop on Mental Models. 25-26 March, Robinson College, Cambridge, 94-102.

GRANT. A. S. and MAYES. J. T. 1991. Cognitive task analysis? In G. R. S. Weir and J. L. Alty (eds) HCI and Complex Systems (Academic Press, London). chapter 6.

HOLLNAGEL, E. 1992a, Coping. coupling and control: the modelling of muddling through. in Proceedings of the Second Interdisciplinary Workshop on Mental Models, 25-26 March. Robinson College. Cambridge, 61-73.

HOLLNAGEL. E. 1993, Reliability of Cognition: Foundations of Human Reliability Analysis (London, Academic Press).

LEPLAT, J. 1989, Cognitive skills at work, in L. Bainbridge and S. A. Ruiz Quintanilla (eds) Developing Skills with Information Technology (Wiley. Chichester). 35-63.

PEW, R. W., MILLER, D. C. and FEEHER. C. E. 1981, Evaluation of Proposed Control Room Improvements through Analysis of Critical Operator Decisions. Electric Power Research Institute, Palo Alto), CA, Report NP-1982.

RASMUSSEN, J. 1980, The human as a systems component. in H. T. Smith and T. R. G. Green (eds) Human Interaction with Computers (Academic Press, London).

RASMUSSEN. J. I993. Analysis of tasks, activities, and work in the field and in laboratories. Le Travail Humain..

RASMUSSEN, J. and JENSEN. Aa. 1974. Mental procedures in real-life tasks: a case-study of electronic trouble shooting. Ergonomics, 17, 293-307.

REINARTZ, S. J. 1989, Analysis of team behaviour during simulated nuclear plant incidents. in E. D. Megaw (ed.) Contemporary Ergonomics 1989 (Taylor and Francis. London) 188 193.

REINARTZ, S. J. and REINARTZ, G. 1989, Verbal communication in collective control of simulated nuclear power plant incidents. in Proceedings of the Second European Meeting on Cognitive Science Approaches to Process Control. 25-27 October. Siena. Italy. 195-203.

RUMELHART. D. F. 1977, Towards an interactive model of reading. in S. Dornic (ed.) Attention and Performance V (Erlbaum. Hillsdale NJ).

WINSTON, M. E.. CHAFFIN. R. and HERRMANN. D. 1987. A taxonomy of part-whole relations. Cognitive Science, 11, 417-444.

©1998 Lisanne Bainbridge