Some of the figures are poor quality scans, available on request.

Lisanne Bainbridge

1. Introduction

This paper discusses the use of working storage and its place in the organisation of cognitive behaviour, and introduces a technique for describing these explicitly.

These issues arose during a study of verbal reports (protocols) from subjects doing a process control task. In such a task the operator (subject) controls a (manufacturing) process: he maintains it in a steady-state by continuously checking for errors and making corrective actions if necessary. The data below come from a task of regulating the overall use of electric power by five steel-making furnaces. Each of these furnaces goes through a sequence of stages in the steel-making process, and each of these stages requires a different amount of power. The subjects controlled a digital simulation of this process, and were asked to 'think aloud' while doing the task. These verbal reports were recorded. These reports showed that the controller alternated between checking for a control error and other types of behaviour. When the process was in an unacceptable state the controller chose which furnace to change the power supply to, and the size of the change. When the process was under control he monitored the present and predicted the future stages of all the furnaces. He also assessed whether any future changes in furnace operation would lead to an unacceptable control state, and if so what the correction should be. The complete analysis is presented in Bainbridge (1972), a summary in Bainbridge (1974).

This was not a problem-solving task, as the controller worked within pre-defined aims, strategies and priorities. The task did not develop from a starting point to a solution, the same behaviour was required repeatedly to maintain a constant situation. The subject was interacting with and controlling a dynamic environment. Two aspects in particular of the verbal protocol behaviour were difficult to model, its flexibility and its parsimony. Behaviour with the same general aim varied in detail or method on different occasions; for example the size of control action to make might be chosen by numerical calculation (digital) or by judgement (analogue). There were no unnecessary repetitions of previous behaviour; for example, when a control error was developing the controller would choose an action, when a later check of the control state showed that the error was fully developed he made this action without repeating the decisions. The verbal protocols were analysed by the usual method of attempting to develop programmes which would produce the same sequence of behaviour as in the protocols. It was found that both the flexibility and parsimony of behaviour could be modelled by making the use of working storage in these programmes explicit, using the technique described below.

This technique was originally developed to account for particular data. Now it has been developed however it may be useful for describing cognitive behaviour in other tasks, particularly for specifying working storage requirements. We can also look at the general properties of the technique. The description of the power control task which was made using this technique has several features which may have general implications for the nature and organisation of cognitive behaviour.

Section 2 introduces the method for explicitly including working storage in descriptions of information processing routines, and discusses when to use stored data, the choice between alternative methods of obtaining the same data, and the relations between routines. Section 3 describes the determinants of the sequence in which routines are used, the place of different types of working storage in different levels of cognitive organisation. and the nature of context effects. Section 4 questions the nature of the limits to working storage capacity, and Section 5 discusses the relation of working storage to longer-term learning.

2. Storage and routines

When studying information processing in cognitive tasks one wants to know what operations are carried out and how these are executed. including the use of working storage. The technique used to describe the information processing must make these explicit.

An example of the simplest traditional method for describing processing routines is given in Figure la. This type of description implies a rigid blind mechanism. The sequence of operations is always the same when the routine is used. The aim of the routine, the data required (in this example the overall power usage after an event) is not named until the end of the routine, at the operation by which it is found (unless the routine has been developed as a subroutine or procedure). This technique is inadequate for the present purpose.

Figure 1 (13K scan) : Routine to calculate overall power usage after change in process stage :
a. described using traditional method.
b. 'box' description of same routine. Many operations which were implicit in (a) have been made explicit here.
c. box symbol in a conditional statement.

2.1. Explicit working storage

The descriptive technique used in the power control study is illustrated in Figure lb. The routine is centred round the data item required, which is named at its 'head'. This data item is represented in three parts: the item's name, a location or 'box' for storing its present value, and a link to the routine(s) by which this value can be found. This routine is itself specified in terms of the data items required. These items are also represented in three parts, their values are found by routines, and so on. The way in which operations or routines at one level are actually implemented by operations of a similar nature but at a lower level is therefore made explicit.

Distinguishing the data item from the routine gives the classic independence of a routine from the point at which it is used, the separation of instrument from purpose. This is useful for several reasons. The same sequence of operations may be used for several purposes, for obtaining similar data items in different parts of the task : see for example the use of the 'choice of furnace to alter' routine (Figure 4, 9K scan, part b) in Figure 3, 6K scan and Figure 4, part a. Less space will be needed to store cognitive routines if such a sequence of operations is stored once in a way which can be referred to from several points, rather than permanently reproducing it in full in each context where it is needed. If purpose and routine are independent this allows one routine to be called on from several data value storage points; conversely, one of these points could call on several alternative methods for finding the same value. This could provide a mechanism for the flexibility found in the protocol behaviour (see Section 2.2.2.).

The boxes in the diagram make the use of working storage explicit. If a data value has been found and stored earlier in the behaviour, then there is no need to repeat the routine to find this value when it is needed again. An arrow in a box in the diagram shows this back reference to a previous value, and so implies a data storage requirement (see Section 3.2.1 and Section 4)

The first time a data item is needed there will be no value in the box, so all the routines will be carried out and the values stored. The next time this item is needed one could recap the stored result without repeating the routines. This provides a mechanism for parsimonious behaviour without unnecessary repetition, but such a simple mechanism raises several problems. In any realistic behaviour the decision to compute data must depend not only on the presence or absence of a stored value but also on its reliability and relevance (see Section 2.2.1 and Section 3.2.1.). The continuity of working storage however does provide a mechanism for context effects and the way in which these determine the sequence of behaviour (see Section 3.2.2). In the remainder of this section the different functions of the three parts of the data item will be discussed more fully.

2.1.1. Data names. As mentioned above, routines are specified in terms of the data items required. This gives a goal- or data-oriented description, as the aim of a routine is stated before the routine is used. This reflects behaviour in the verbal protocols, the subject often mentioned the data required before using the routine to find it, e.g. 'now, how long has it been melting?'. These are information-processing rather than feedback-control goals. One might call this feedback-control with 'lack of data' as the 'error', but this would trivialise the term feedback, as no quantitatively varying error correction is involved here.

This naming could also describe the way in which subjects (particularly the inexperienced) can know what item they need the value of, but not how to find this value. This situation could be represented by an item without an associated routine. The lack of routine could initiate a search for method, by asking questions, reading task instructions or working it out for oneself. Once a routine has been acquired it could be called on when needed.

2.1.2. Data storage. A 'box' in the diagrams represents the functional location at which data values are stored or referred to.

In the power control study various types of data value were stored. Numerical values could be obtained by : reading displays, written notes or the task instructions; from memorised constants; by calculation using a slide-rule or longhand on paper; or by judgement. Names, for example reference letters, the names of process stages, or quantitative categories (above, alright, action required, etc.) could be obtained by reading displays or by information processing. Some boxes were needed simply to note whether a named item existed or condition was met, e.g. the next power usage changing event compensates for present power usage error - yes or no.

Most of these stored values were data items which were not displayed explicitly to the controller. There was good evidence that some of the displayed values were also stored however and that the controller acted on these stored values, e.g. the subjects sometimes acted according to past rather than present values of the process. Presumably these values were stored because this allows quicker access, or because the stored data was organised in a more useful way than the displays, e.g. the present furnace stages were displayed in alphabetic order, the behaviour of experienced controllers showed that they remembered them in lists of furnaces in the same stage.

Each location stores one value. A series of 'boxes' is used for a list of values. Past, present and predicted future values of the same item are stored in separate boxes. This is appropriate as a different routine is used to find the value in each case. A box containing an arrow acts as a 'token' for data stored elsewhere. The number of items which must be available for reference at any one time can be counted from the diagrams drawn for the power control task, and imply the amount of working storage which is necessary for this behaviour at a particular time (see Section 4).

The word 'box' has been used with quotation marks because it is not intended to imply a particular mechanism for working storage. It has been used as a symbol in the diagram because it is easily identified, and it can be used for noting data values when working through the routines by hand simulation. If working storage has a box-type mechanism some representation of the data value would be transferred to and 'put in' this location. In an alternative mechanism a pointer at this functional storage location would refer to some data value representation which was stored elsewhere. A copy of this representation would be transferred only when its value was actually needed in computation (see Section 5).

2.1.3. Routines. The data name and storage location refer to a routine for finding the required value. Each routine must contain at least one operation which finds this value (symbolised in the diagrams by a double-lined box). Two types of routine were needed to describe behaviour in the power control task. In one the data value was found in the final operation of the routine. In the other type the routine progressively refined the data value by considering more dimensions of the situation, so the routine included a succession of operations which gave a data value, each of increasing accuracy.

Each routine is implemented by routines themselves consisting of three-part items, and so on. Routines also contain other types of component. The routines in the power control task sometimes contained alternative behaviours which were appropriate in different circumstances, and conditional statements were needed to describe the decision between these alternatives (see Section 3.1). The routines did not go to an infinite depth of routines implementing routines, they eventually referred to 'basic operations', indicated by underlining in Figure 1, 13K scan, part b. These might be mental actions. i.e. basic cognitive processes or information 'primitives', or they might be physical actions, interactions with the external world. Some of these physical interactions (e.g. reading, asking questions, using a slide-rule) gave data for use in cognitive processes, so affected the state of the internal world. Other physical actions affected the state of the external world (e.g. by making a control action). These basic operations are integrated into meaningful sequences of activity via the three-part data items, but conversely routines only become operational via these basic operations, in a sense they are otherwise only tokens or symbols.

2 .2. Deciding to use a routine

Several decisions must be made at the point at which the data requirements of one routine lead to the use of another routine. The subject must decide whether to repeat a routine even if a value is already stored. When alternative methods are available he must decide which to use. Decisions at this point could lead to a wide range of different behaviour sequences from the same substrate of available routines.

Any breakdown in these decisions could lead to the sort of errors in working storage usage which tend to occur under stress or fatigue (e.g. Davis, 1948; Welford, Brown and Gabb, 1950). In these circumstances subjects tend to forget sections of behaviour altogether, or to repeat data-finding behaviour unnecessarily, or to act on outdated information. The point at which the subject decides whether to use a routine could be the locus for these errors.

2.2.1. Deciding to repeat a routine
The subject should decide to repeat a routine if the stored data value is either untrustworthy or irrelevant. For example a stored value may have decayed or have been disturbed (if the limit to working storage capacity is such that the stored values are continually changing), so that the value is no longer a true record of what was originally stored. The external world may have changed dynamically with time, so the stored value no longer represents the actual state of affairs. Finally, if the routine was last used for a different purpose its stored value will be irrelevant.

To be able to decide whether to recompute, the subject must have data about these dimensions of the stored values, and a mechanism for assessing the data. The subject must be able to assess the reliability of a data value as a true record, perhaps from its 'noisiness'. He must also be able to assess the likelihood that a particular dimension of the external world has changed. One could suggest many methods of making this assessment for particular data items in the power control task (it is interesting that many of these methods would require some knowledge of when the last sample was made). But this would be speculation. Behavioural data should indicate the actual processes involved. There is some evidence on the repetition of routines in the power control task, but the data analysed on this so far is rather limited and only simple inferences can be made. When the control error was unacceptable the controller usually repeated all the routines. It is important to be correct in this situation. When control error was acceptable the controller monitored the present and future stages of all the furnaces. He repeated this survey twice, and then started to chat in general, keeping an eye on the process for any change. This might suggest that two repetitions were sufficient for retention of the data values. Unfortunately the idiosyncracies of the particular situations in which this behaviour occurred meant that later behaviour did not need to make use of these data values so did not show whether or not the values were remembered. Stored data values are irrelevant if a routine was previously used for a different purpose. It will be easier to understand the wider implications of data and discussion on this after some more points have been made (see Section 3.2.1).

2.2.2. Choosing between alternative routines
Many data items may call on the same routine and one data item may refer to several alternative routines. This suggests that the item-routine link is a complex one.

In the power control task there were two different types of alternative method for finding a data value. One routine might refine the data value it found, by progressively considering more dimensions of the situation. In this type it was necessary to decide how far to go in adding refinements within one basic method.

In the more usual case there were different methods of finding the same value, as in Figure 2. so it was necessary to decide between them. Both these types of choice between alternatives could lead to a speed-accuracy tradeoff in performance, as the less time allowed the less accurate the method used.

Figure 2 : Suggested nature of link between data item and alternative routines for finding its value. Each working method has stored with it information about what would be involved in using it, which is used in deciding what to do.

Data from the power control task suggest that the factors used in deciding between alternative methods may be : the difficulty of the methods, the time they take, and the amount of working storage they need. The controller found the value of the control error by judgement instead of making long complex calculations (data in Bainbridge, 1971). He alternated between action choice routines and checking the control error, when error was unacceptable. When the control error was more unacceptable the interval between control error samples was shorter, implying that shorter routines were used. This suggests that the urgency of the situation influenced the duration of the routine used. The 'choice of furnace for action' routine in Figure 3, 6K scan and Figure 4, 9K scan illustrates the type of data which might be obtained on the effects of working capacity requirements. This routine was used at two different 'depths' : alone (direct from Figure 3), and via the 'future assessment' routine (Figure 4a, which is called by Figure 3). If there is a limit to working storage capacity then there might be less capacity available for use by a routine when it is called via another routine, so the routine might be simpler in form. The behaviour shows that this choice of action routine was simpler when used in choosing a future action. Unfortunately this particular example does not provide good evidence about working storage capacity, as possible alternative actions are fewer in the future than in the present, so the routine for choosing a future action can be simpler. We are only beginning to understand how to find evidence on this type of question.

The data suggest that the above factors influence the nature of the routine used. This suggests that the following are required by a data item, to provide a link to the routines. There must be a list of the alternative routines (with pointers to where they can be found). Data on the costs and values of using each routine must be available, at least relative measures of the difficulty of the cognitive processes involved, the time and working storage needed, and the value (e.g. the accuracy obtained) of using it. There must also be a decision mechanism which can assess the currently permissible levels of these costs and values, and use them as criteria in choosing between the routines.

2.3. Levels in the organisation of routines.

It is traditional to describe the organisation of behaviour in terms of the branching levels of a goal-oriented hierarchy (cf. Bryan and Harter, 1899; Miller, Galanter and Pribram, 1960; Newell and Simon, 1961). In practice however it is often difficult to define what constitutes a 'level', or to identify the level of a particular piece of behaviour. The present technique describes behaviour as a goal-oriented branching structure. The organisation here is not hierarchical however, and the reason for this demonstrates both why it is difficult to identify the levels in any real behaviour and also the general inadequacy of a hierarchy as a description of behaviour.

The key to these limitations lies in the lack of unique links between goals (data items) and the routines for implementing them. These links may be unique because of limits to the logical or physical possibilities in a particular task, or because a subject has not learned that some routine has more general application, but this is not necessarily so. The result is that it is not possible to draw a unique or 'tree' type of description for the branching in task organisation.

Each data item calls on routines which consist of data items, and so on. This apparently defines level of organisation. It does not define a constant value of level for any particular routine however. If items are free to call on any routine then a particular routine may be called on by any of these levels, so that its own level can only be defined in relation to a particular context, not in general. (See the 'choice of furnace for action' routine in Figure 3, 6K scan and Figure 4, 9K scan.) It is also at least logically possible for one routine to lead to another and so on until it leads back to itself. In such a freely interconnecting network one item may at a particular time be the 'top' item from which others lead. If other data items can also be the focus of organisation however, then the level of a particular routine will not be constant but will depend on its relation to the top item at any particular time. Thus levels cannot be identified within the static structure of task organisation, but a functional hierarchy may appear when this structure is used dynamically to do a particular task.

Complete freedom of interconnection between routines does not appear in real behaviour. This freedom will be limited by the possible meaningful relations in the task. Also any limit to working storage capacity may restrict the number of routines which can lead on from one another at a particular time. In the power control task a 'top' and 'bottom' to the interrelations between the routines were defined by the sequencings 'above' the routines, which determine the order in which the routines are used (see next Section), and the basic operations 'below'. Within these limitations one can suggest however that routines are related in a network or heterarchy of interconnections, rather than in a hierarchical tree.

3. Conditional determinants and behaviour sequencing

In the power control task the verbal protocol could be divided into sections which were described by about a dozen routines. These routines found the values of a dozen main data items. These were identified as main or 'top' data items as they were not necessarily named within another routine when they occurred, finding their value could be the overall aim in a section of the protocol. Examples of these items are the furnace to alter, the size of the control error, the size of action to make, or the effect of a future event.

Once these basic units of the behaviour had been distinguished, the variables determining the sequence in which they occurred could be identified. These determinant variables, and their effects on the sequence of behaviour, were then described by flow diagrams which consisted chiefly of conditional statements, but also contained data items and physical actions. The conditional statements described the (decisions between) alternative behaviours. These 'sequencing' diagrams were not subordinate to some head data item, to finding the value of a single variable. Any single overall name for a sequencing would have to be the name of a type of activity, not the name of a data item. About five different sequencing diagrams were developed, which determined the routines to use when the control error was acceptable, unacceptable etc. Figure 3 is a simplified version of the sequencing when control error was unacceptable.

Figure 3 ( 6K scan) : Simplified description of sequencing when control error was unacceptable. In this and later figures, routines may be described verbally or using brief implicit notation, for simplicity.

Figure 4 (9K scan) : Routines used by sequencing in Figure 3 :
a. assessing effect of future event.
b. choice of furnace for power supply change. This is used in both Figures 3 and 4a.

In all, there are three distinct types of 'branching' in the description of behaviour used here. There is one type of branching in the way that one routine is implemented by others. Branching exists where there are alternative methods for finding one data item. The choice between alternative behaviours which are appropriate in different circumstances is a third type.

3.1. Conditional statements

The choice between alternative behaviours is determined by comparing a variable value with a criterion e.g. if control error is greater than +3 units the controller acts, otherwise he monitors. This choice is usually described in a programme by using a conditional statement containing the variable and criterion. In the power control task conditional statements appear in the routines and more particularly in the sequencings.

We can consider whether it would be useful to describe the variables in these conditional statements in the same three-part way as used above, distinguishing between name, storage location and link to routine. At points in the power control protocols where alternative behaviours were available, the data showed that different methods might be used to find the value of the variable determining which behaviour occurred. Also, if a value of this variable was already available the routine to find it was not repeated, very often the appropriate behaviour followed immediately without explicit mention of the determinant, as if the behaviour followed a preset switch. These findings suggest that the same methods should be used, for the same reasons. An example of this type of conditional statement is given in Figure 1, 13K scan, part c. The data storage location is incorporated within the decision test, to imply a mechanism which switches to the next appropriate behaviour when the data value is available.

This mechanism can be extended to describe cases where the behaviour sequence is determined by multiple conditions. There are two ways of representing the effect of multiple conditions, these methods are equivalent in outcome but not in mechanism. One method uses a branching tree of (not necessarily binary) conditional statements. The other uses table-look-up; all the determinant variable values are used simultaneously to define a cell in the table, the behaviour to do next is named in this cell. We can consider which of these methods would best describe the cognitive behaviour, given the evidence from the power control task. Branching conditionals imply that the determinants have their effect via a sequence of one-dimensional decisions. In the power control task, when all the determinant data values required were available the determinant variables were not mentioned and the appropriate behaviour appeared rapidly; this would support a table-look-up mechanism. However, if the determinant values were not available then routines to find them must be executed in sequence. A special mechanism for this would have to be added to the table-look-up method. These data-finding operations are inherent in the conditionals method, while if the data values are available then each condition can act as a switch. A sequence of conditionals could behave like a vector of switches, and so have an immediate effect. It might appear difficult to describe interactions between determinants by using branching conditionals. There is no evidence of this in the power task, but if interactions are dealt with by interaction in cognitive processing then this could be incorporated.

3.2. Functionally distinct levels of organisation

The sequencings and the routines used to describe the power control behaviour differed in that the sequencings were not subordinate to some head data item, sequencings determined the overall use of the routines, and all operating instructions were within the routines. This suggests that we can identify sequences and routines as two structurally and functionally distinct types of organisation. Sequencing is a level of organisation 'above' the routines, as the sequencings determine how the routines are used. In the power control behaviour, sequencings could follow on from sequencings, in a way which showed that the relations within the sequencings level of organisation could also be best described as a network rather than a tree.

The nature of the link between the sequencings and the routines is particularly interesting. The set of main data items and the set of sequencing determinants in the power control task were identified independently using different methods. Importantly, it appears that the main data items and the sequencing determinants are very similar (see Table in Bainbridge, 1974). We can therefore suggest that the main data items are the items which occur in the sequencings, particularly within the conditional statements. The 'top' of the routines level of organisation (the main data names, storage locations and pointers to main routines) is within the sequences level of organisation.

Two further levels of organisation could be identified in the power control task, making four levels in all. The basic operations formed a distinct level 'below' the routines. The order in which the sequencings occurred was determined by a higher 'overall' level of determinant variables, particularly by the size of the control error, as described in Figure 5. In the power control task it was at this level that the controller judged the adequacy of control and selected the appropriate general type of behaviour, at this level the behaviour interacted with the feedback control task in the external world.

Figure 5 : Overall determinants of sequencing. Depending on the nature of the control error, the next activity may be one of three types.

3.2.1. Level of organisation and working storage. As the names and storage locations of the main data items are actually within the sequencings, the use of working storage might differ with level of organisation, and evidence from the power control task does suggest this.

Two types of working storage were needed in the diagrams developed to describe this task. 'Temporary' storage was required for any data values found in a routine which were used again later in the same routine. A single arrow in a box indicates back reference to such a value. Longer term 'cross referenced' storage was needed for data values which were found during one routine and made use of during another (shown by a double-lined arrow). Inspection of the diagrams for the power control task gives evidence about the incidence of these two types of storage. This is only weak data about storage, as it is extracted from descriptions of the behaviour, not from the behaviour itself.

All the temporary storage occurred within the routines, and so within the routines level of organisation. (In fact this is necessarily so, as sequencings do not contain instructions for computations, etc.) Each temporary item needed to be stored for a few seconds at most. The next items required temporarily might be completely different in nature, e.g. times then furnace names, etc.

Most cross-references in the diagrams were from within the sequencings level, either in conditional statements or as parameters of routines finding the main data items. All the main data items were cross-referenced or made use of at some other point in the cognitive behaviour. Importantly, all the cross referenced data items were in the sequencings. This suggests that all this longer term working storage was within this sequencings level of organisation.

Although this relation between level of organisation and durability of storage is an empirical finding, it is actually a necessary characteristic of any information processing organisation in which routines are independent of the purpose for which they are used. In such cases, data values found in one use of a routine are not necessarily relevant the next time the routine is used. Any data values relevant to a particular purpose must be stored separately from the routine to avoid loss or confusion. The question of how to judge whether stored data values are irrelevant so a routine must be repeated was raised in Section 2.2.1. This discussion suggests that all data values below the main items are recomputed, so that decisions about relevance do not arise. This also implies that though the main data values will be remembered, the data used in finding them will not be. It would be interesting to know whether these findings are also true of other cognitive tasks. The power control task was cyclic in nature, and some data values remained the same over a period of time. The same effect might not be found in sequential tasks, such as maintenance, diagnosis, or problem solving.

Stronger evidence about the longer term storage of main data items should be obtained directly from the behavioural data. We require evidence that main routines are not repeated unless the stored value is unreliable or has changed, and evidence that these values are cross-referenced instead of recomputed when the data is needed elsewhere. The protocol data was analysed for such evidence by charting the protocol behaviour when each main routine should have occurred, and noting the incidence of repetitions. Unfortunately the nature of the data makes it difficult to obtain firm evidence for storage. In verbal report data, if a particular operation is not mentioned but its consequences appear, this is not evidence that the consequences depend on storage rather than executing the operation, as the operation may have been done but not mentioned. This difficulty is somewhat reduced in the present data, as there was evidence that this protocol reported a random sample of the operations carried out, so one might infer at least that if an operation never appears in a given context, it is never done. Against this, it is unlikely that a routine will never be repeated, as the stored data may be unreliable or unrepresentative, so that such evidence is unlikely. Really strong evidence for storage is only obtained when the subject acts on a past rather than present value of an item. Despite these difficulties there was reasonably good support at least for retention of the results from decisions about the size of the control actions, and for the subject keeping his place within the general survey of all furnace states.

No comments can be made here about any use of working storage in the basic operations level, as these operations are not verbalised. At the 'overall' level the main data item was the size of the control error. This was frequently recomputed of course, as it could change rapidly. There was strong evidence however that its value was retained between computations. It was an important determinant referred to by many conditional statements in the sequencings (see e.g. Figure 3, 6K scan and Figure 4, 9K scan). It was also a major criterion the decisions which select which routine to use, according to the urgency of the situation (see Section 2.2.2).

3.2.2. Context effects and prediction of the environment. The decisions which determine the sequence of behaviour refer to the values of the main data items. These item values can be identified as both the 'context' in which the task is done, and the controller's 'mental picture' of the present state of the process. In this task, the storage of these values in the sequencing level of organisation provides the location and mechanism for these concepts and their effect on behaviour. As these data values are obtained by working through the data-finding routines, so this general orientation to the task takes time to develop.

Some of the main routines made predictions, such as what the next power changing event will be, when it will occur, what its effect on control error will be, and what action will be needed when it occurs. As data items these predictions did not differ in type from the results of other routines, they could also affect the sequence of behaviour. For instance, there was no need to make an action to reduce control error if an imminent furnace state change would change power usage by the required amount, see Figure 3, 6K scan and Figure 4, 9K scan. Thus predictions were also part of the context or mental picture of the environment, and the importance of this was that external events were anticipated.

In the power control task the controller's ability to make and use predictions developed with experience Inexperienced controllers (university students) at first used feedback control, changes in the process were unexpected and dealt with after they occurred. More experienced controllers (steel workers) predicted and anticipated these changes. This was a form of feedforward control which allows much greater control over the environment.

3.3. Practical uses of this technique

A technique which describes information processing tasks in a way which makes the use of routines and working storage explicit can have many uses. This technique might be used in experimental studies of working storage to specify a priori the amount of working storage required in a given task at a particular time. This could give a measure of the relative difficulty of using working storage in different tasks. It could also provide a normative basis for developing tasks of different levels of difficulty.

A task description in this form might also provide a job-aid for human factors and training scheme designers. It could help the human factors designer to identify the items which are used together so should be placed together on a console. It also indicates variables which are not directly monitored from the process but could usefully be computed and displayed. The power control task data suggest a warning about this however. In order to make rapid decisions the controller needs an up-to-date mental picture of the state of the environment. This does contain data values which are explicitly displayed, and it takes time to develop. This may account for the way in which industrial process controllers take time to get the feel of the system when they come on shift. It also suggests that if rapid manual-takeover of control is required when there is a breakdown in automatic control, then the controller must have been actively involved in the ongoing events and decisions of the process, even though this is not technically necessary, so that his mental picture of the process is fully developed and up-to-date and can be used immediately for appropriate control decisions.

This technique might also provide reminders about various aspects of training scheme design. It indicates which routines and basic operations must be learned, and their interrelations. It should make explicit any alternative methods available, and the costs and values of using them; some of these methods may need to be encouraged or discouraged in training. It also shows what working storage is required, particularly the items found in one part of the task which are used in another. The data from the power control task suggest that inexperienced subjects take some time to appreciate the possibilities for cross-referencing data, if left to pick them up for themselves. It is necessary to point out however that this technique should probably only be used as a basis for training of routine tasks. It does not necessarily describe the task operations in such a way that the controller knows how they can be recombined to deal with new situations.

4. Working storage capacity

The discussion so far has concentrated on the location and use of working storage in information processing. Any complete account of this storage must also discuss its capacity, and the nature of capacity limits. Inferences about the storage capacity required in the power control task can be made from the diagrams describing the behaviour.

The items which must have been stored during the power control task were the items referred to by the boxes containing arrows. The number of items which must be stored at one time varied dynamically throughout the behaviour, as some items were no longer needed and others became necessary. It is therefore most useful to count the maximum number of items which may be required simultaneously. The controller made written notes of some data values, particularly of arbitrary values such as event times, so storage of these was not necessary. A count of storage requirements which does not allow for this may be misleading. (Newell and Simon (1972) have initiated an interesting discussion on the effect of this 'external memory' on task strategy.)

In the power control task a maximum of four temporary' items might be required simultaneously by a routine. The majority of the sequencings used a maximum of three 'cross-referenced' items, but the sequence which surveyed the present and future states of the furnaces involved up to 19 such items. As the temporary and cross-referenced items were required at the same time these figures imply that a capacity fop storing seven to 23 items was necessary to do this task. Experimental data on short-term storage indicates a capacity of about seven static items, e.g. Miller, 1956, or two to three running items (Yntema and Meuser, 1960, Bisseret, 1971). The capacity requirements of the power control task were apparently larger than this, so raise several questions. Either the working storage capacity can be larger than indicated by the above experiments, or less working storage was used in the power control task than is suggested by counts from the diagrams. There might have been restrictions to the use of working storage which have not been identified in the present study. For example only subsets of the cross-referenced items might actually be available at a particular time.

Several points suggest that storage capacity might be larger than is indicated by the above experiments. The apparent capacity may be a function of the technique used for measuring it. In the same way that the apparent amount of material retained in long-term memory depends on whether this is measured by recall or recognition. The above experiments were all strong tests of memory, as they used recall. Also this was recall of arbitrary items. Different results might be obtained by using meaningful material in a patterned context. The possible variable values in the power control task came from a limited set. For example there were only eight different furnace stages: the different sequences in which these could occur were very constrained, variations in the sequence depended on the size of the furnace and the type of steel being made. Once such constraints are known, at any one time the rest of the task will give many recognition cues to the value of a particular item. The classic experiment on recoding quoted by Miller, op cit, apparently showed that the effect of patterned organisation on the amount remembered is such that capacity is still limited by the number of items. The untrained subject remembered a certain number of binary items: after training on more complex codes the subject could remember the same number of items or 'chunks', each of which had a higher information content. Although the notion of a chunk is useful in thinking about the organisation or patterning of material, it is difficult in practice to find single names to represent each complex organisation or chunk, except where there is a 1:1 mapping between different entropy codes as in Miller's experiment. It may be that the important aspect of that experiment is the demonstration that learned patterns of organization allow a subject to handle a larger quantity of material, and that the fact that the subject remembered the same number of items after learning in this experiment is a bit of a red-herring.

Obviously a great deal of further evidence is needed to investigate and resolve these questions. The actual behavioural data. rather than the diagrams, which have been analysed so far for the power control task contain too few examples to give more than anecdotal evidence. Some types of protocol behaviour would show that data values are not available from storage. For example, if the controller worked through all the routines after changing from one sequencing to another, this would imply that no relevant cross-referenced data was available. If routines were simpler in form when used deeper in behaviour this would imply a capacity limit (see Section 2.2.2). (Conversely, the behaviour would indicate that stored data was available if the controller could go through a sequencing without repeating the routines. One section of the observed behaviour indicates the possibility of a large capacity : the controller alternated, apparently independently, between considering a power increasing action to correct a present low power usage, and considering a future power decreasing action to correct the effect of a future furnace change when power usage will be too high. Storage would also be indicated if the controller's behaviour was appropriate to a past rather than present state of the process.

The discussion so far has been concerned with the use of working storage within a task. Storage is also necessary for the controller to be able to do other types of observed behaviour: to return to his place in the task after an interruption; to do more than one task at a time (storage is needed both in scheduling attention between the tasks and in keeping track of one while doing the other(s)); and to note that new data accounts for some unexplained past event. These uses of working storage are not necessarily made explicit by the present technique.

Evidence from the power control task has questioned the limits to working storage capacity. This descriptive technique might give some clues to the nature of the capacity limit. Whatever the mechanism of storage it must be very flexible to allow for the rapid changes in temporarily stored items. Using pointers would be a simple way of implementing this flexibility. Temporary items might not be retained simply because they were not referred to again. Cross-referenced items might be rehearsed by the process of reference. The above discussion on the possible effects of meaningful context suggests that capacity may also depend on much more complex mechanisms, such as pattern recognition, which are by no means accounted for in the present technique. The next section will discuss this further, and will raise other points which show that any simple notion of working storage is inadequate.

5. Working storage and learning

The word storage has been used throughout this paper, rather than the word 'memory'. Traditionally memory involves the ability to reproduce past events, a function which is not accounted for by the mechanism described here. 'Storage' refers to data values which describe the present and future states of the environment and which are required temporarily while doing a task. These values originate from samples of the external world, for inexperienced subjects at least. Although these stored values reflect the variety in the environment and decisions about it they can still provide the basis for learning.

Learning the values of constants is a simple case. Repeated sampling of a constant-valued data item will give the same result each time, so the pointer from the storage location to the data value representation can become permanent. The dynamic character of short-term storage can change to the static structural nature of long-term storage simply because the stability of storage reflects the stability of the environment.

When a data item has alternative possible values the case is more complex. For instance in the power control task there are five different furnaces, and eight different furnace stages. The controller can soon learn this set of alternatives, in a simple case. One can suggest that the storage location then has access to this set of alternatives, and it is only necessary to identify which one is relevant at a particular time. This could have various effects on the efficiency of performance. In choice reaction-time experiments, practice reduces the amount of time required to handle a given number of alternatives (Mowbray and Rhoades (1959)). A reduction in handling time per item would have implications for rehearsal models of storage capacity limits. In this type of theory, capacity is limited by the number of items which can be rehearsed in a given time. If each item comes to require less time then more items could be rehearsed within the time limit, so capacity should increase with practice. The reduction in time presumably reflects a change in the difficulty of handling the data values. This might occur for instance because the set of pattern recognition devices required have been identified, or developed and refined.

For some data items there is a large number of alternative values. and these are not peculiar to the particular task. For example 'time' (e.g. 14.56) uses numbers, which are also used for other purposes. In a simple pointer-to-representation mechanism for storing time values, all possible times would have to be represented. This is obviously uneconomical. Of course representations of all the digits are required, but for longer numbers it would be more parsimonious to have an extra mechanism which specifies the sequence of digits to store at a particular time. Laboratory studies of memory have invariably used these non-task-specific and infinite-set types of item, which may require special mechanisms to handle.

For many data items the possible values are not discrete alternatives but vary along a continuum. Examples in the power control task are the level of power usage and the size of the control error. Different samples of these data items can give infinitely graded different values, but at least these values can only vary within the dynamic limits of the process (environment). Each new sample must therefore both fit in with and contribute to the controller's knowledge of the possible states of the process. With increasing knowledge of the process dynamics the controller can make more accurate predictions (judgements) about the effects of actions or events on future states.

This type of dynamic knowledge might be represented and accessed either by an input/output table, or by a general template or pattern of the dynamic changes. An input/output table, using the mechanisms available in the present technique, could be entered by using multiple-output conditionals. These could not reasonably distinguish between all the values on a continuum, so would have to be restricted to distinguishing a limited number of categories of value. This mechanism therefore implies that absolute judgement or category scaling would be used to measure values on the continuum. There is good evidence that this is so in process control (e.g.., Cooke, 1965, Bainbridge, 1971). A general template of the possible dynamic changes could be used by a routine to generate the relevant state in particular circumstances. Such a routine could be included with the other routines used in doing the task.

The flow diagrams or programmes for the power control task represent the controller's knowledge of appropriate information processing operations. This can be distinguished from his knowledge of the static and dynamic characteristics of the furnace process. The latter knowledge could be identified as the controller's 'mental model' of the external world. The values in this knowledge could be the data value representations referred to by the pointers from the storage locations in the processing programmes. The working storage data values, or 'mental picture', would be the subset of these possible values which actually exist at a particular time. The pointers from the storage locations to these values would be the links between the programmes and the mental model.

It may be however, that the distinction between programme and data is not so clear as this implies. This distinction may be an artefact of the technique used to describe cognitive activity here. A minimum number of routines and sequencings were developed to account for the consistent repeated sections of power control behaviour. There are still similarities however between what have been described as distinct routines and between them and sections of protocol which do not match any of these routines. This might imply that the controller has a general structure of knowledge about the process and the operations which it and he can carry out, from which sequences of behaviour can be generated for particular purposes. The consistently repeated sequences of behaviour which have been identified as routines might appear because the same circumstances recur and the response to them becomes overlearned. Whatever form this structure of task knowledge takes it must allow its possessor to overview the static structure and dynamic functioning of the external world, to describe as well as to perform his task operations, and to comment on and revise his performance and strategy. It must also incorporate working storage as a basis for learning about both data and operations.

A.Bisseret has pointed out that while the air traffic controller remembers about three attributes of an aircraft he remembers these values for 9-10 aircraft, his measures are therefore similar to the capacity reported here.

References

BAINBRIDGE. L. 1971. The influence of display type on decision making. I.E.E. Conference Publication No. 80. Displays.

BAINBRIDGE. L.. 1972. An analysis of a verbal protocol from a process control task. Unpublished Ph.D. Thesis. University of Bristol.

BAINBRIDGE. L., 1974, Analysis of verbal protocols from a process control task. In: Edwards and Lees (Eds.). The Human Operator in Process Control. Taylor and Francis.

BISSERET. A.. 1971. Memoire operationnelle et structure du travail. Bulletin de Psychologie, 1970-1971, 24, 280-294. English summary in Ergonomics, 1971, 14, 565-570.

BRYAN, W.L., and HARTER. N.. 1899, Studies in the telegraphic language. The acquisition of a hierarchy of habits. Psychological Review, 6, 345-375).

COOkE. J.E., 1965. Human decisions in the control of a slow response system. Unpublished D.Phil. Thesis. University of Oxford.

DAVIS, D.R.. 1948. Pilot Error. Air Ministry Publication A.P. 2139A., H.M.S.O.

MILLER, G.A.. 1956. The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychological Review, 63, 81-97.

MILLER, G.A., GALANTER. E. and PRIBRAM. K.H.. 1960. Plans and the structure of behaviour. New York: Henry Holt and Co.

MOWBRAY. G.H., and RHOADES, M.U., 1959. On the reduction of choice-reaction times with practice. Quarterly Journal of Experimental Psychology, 11, 16-23.

NEWELL. A.. and SIMON. H.A., 1963. G.P.S., a program that simulates human thought. In: Billing (Ed.). Lernende Automaten. Oldenbourg. 1961. Reprinted in Feigenbaum and Feldman (Eds.). Computers and Thought. Macgraw-Hill.

NEWELL. A.. and SIMON. H.A.. 1972. Human Problem-Solving (Prentice-Hall).

WELFORD. A.T., BROWN. R.A. and GABB. J.E.. 1950. Two experiments on fatigue as affecting skilled performance civilian aircrew. British Journal of Psychology, 40, 195-211.

YNTEMA. D.B., and MEUSER, G.E.. 1960. Remembering the present states of a number of variables. Journal of Experimental Psychology, 60, 18-22.

©1998 Lisanne Bainbridge

Main Index Page

Comment Form