This is the fifth of nine sections :
1. Introduction
2. The cognitive processing element.
3. Ways of meeting the cognitive needs
4. Sequencers and the contextual overview in working storage.

Apologies for the poor quality of many of the Figures.

Lisanne Bainbridge

The term working 'storage' is used in this paper, rather than working 'memory', because this working storage is not simply a memory for items which are replicas of the input information. Working storage here is a task-related temporary structure of data which has been transformed and interrelated as a result of task thinking to meet the cognitive needs, cp. Figure 2.3.2 and Figure 4.4.1. However Section 5.3 will suggest a possible relation between working storage and working memory. Previous sections have focussed on the cognitive processing element and its use as a building block for more complex processing structures. This section begins to discuss aspects of cognitive behaviour for which additional mechanisms may be needed.

5.1. Updating the data in working storage

From the simplified representation in Figure 4.5.2 it may look as if there is no way for the processing to get started. However, this is dealt with by the two-arrow box described in Section 4, part 4.5. This provides a mechanism for finding information if it is not already available. But a mechanism which only looks for information which has not been obtained before will never update that information once it has been obtained, and will assume that the external world continues in an unchanging state. Some mechanism is needed for keeping up-to-date with the changing state of the external world. In the behaviour of the highly experienced steelworks operator, this appears to take two forms, related to what the operator is updating. To discuss this it is useful to distinguish between two aspects of the state of the external world :

1. its state relative to the main task goals, e.g. in the furnace control task : whether or not the electric power usage is within acceptable limits.

2. the general state of the other parts of the external world which it is relevant to know about when thinking how to meet the main task goals, and which is used in sequencing decisions. These are data used in optimising the response to the main task goals and meeting secondary goals. For example, in the furnace control task : what each furnace is and will be doing, and the effect this activity has on electric power usage, and on the best action to make, if required. Updating these data about the wider situation will be discussed first.

Figure 5.1.1.1 ..(9.2K) : The ONGOING 'sequencer' used when updating awareness of the general control situation, and predicting future events and their effect on the future control state. These cognitive activities are done when the control state is acceptable (see Figure 5.1.2.1). UPDATE is an activity done in several contexts. (Bainbridge, 1972, Figure 7.3.3). (For ASSESS see Figure 5.1.1.2).

5.1.1. Updating the data on the general state

The furnace operator spent his time, when action was not needed (see Figure 5.1.2.1), in reviewing the general state of the process. This behaviour was described in the simulation by a 'sequencer' called ONGOING. This 'sequencer' reviews the present state of the process (Figure 5.1.1.1), and predicts forthcoming events, their effects, and what to do about them (Figure 5.1.1.2). The review of the situation which is built up in this way is then referred to during decisions about what to do next, as indicated in Figure 4.4.1. So there are explicit and distinct cognitive processes, in a 'sequencer', for updating the representation of the present situation. Beishon (1969) also found that the forward plan of his oven operator included explicitly making an overall check of what was happening on the ovens.

Figure 5.1.1.2 ..(9K) : Supplements to Figure 5.1.1.2, activities to assess future state changes predicted as in Figure 5.1.1.1 (Bainbridge, 1972, Figure 7.2.6).

This type of updating leads to a mechanism for one type of cognitive error. If the decisions for choosing what best to think about next have not led recently to doing this updating, then the working storage representation of the state of the task may not be valid. For example, in another study of process operation, an operator turned a control down when it should have been turned up. Without evidence about his previous cognitive activity, this would have appeared completely irrational. However, the operator had previously predicted that when a particular event occurred in the future, it would be necessary to turn the control down. In the period intervening before the event, he had been too busy with other activities to notice that something else had happened which changed the situation so that the preplanned action was no longer appropriate. As is often the case with human cognitive errors, a mechanism which usually increases behavioural efficiency occasionally gets caught out.

5.1.2. Updating the control state

The furnace operator's behaviour, when checking his main task goal, is described in the simulation by the OVERALL 'sequencer', described in Figure 5.1.2.1. This 'sequencer' calls on the main 'routine' for checking the control state (number 1, described in Figure 4.6.2) and chooses the next best behaviour on this basis.

Figure 5.1.2.1 : Choosing the most appropriate next cognitive activity, given the control state as assessed by the 'routine' [1] described in Figure 4.6.2 (Bainbridge, 1972, Figure 7.3.a).

Figure 4.2.3 represents the operator as checking the control state (i.e. returning to the OVERALL 'sequencer') only after working through the whole of another 'sequencer'. However, things are not so simple. The verbal protocols show that the operator may briefly check the control error during thinking about another part of the task, and then, if the control error is unchanged, go on with the other task thinking from the point at which he left it. In a 'routine' this can happen between cognitive needs. For example, in Figure 3.2.2.3, the operator may check the overall control state after checking the power levels, and then go on to looking at furnaces which were charging. In a 'sequencer', the operator may check the control state at the end of a section of processing. For example, in the ONGOING behaviour described by Figure 5.1.1.1, the operator may check the control state at the end of each of the three branches (which can be done in any order), not just after all three have been worked through.

This interrupting suggests that different 'sequencers' have different levels of priority, and that higher priority 'sequencers' can interrupt lower priority ones. This could be important in mechanisms to account for multitasking (see Section 6).

An inverse issue arises. Perhaps the operator does not just check the control error at the points explicitly directed in the 'sequencers' and 'routines', but checks it after meeting every cognitive need in each 'sequencer' and 'routine'. The data suggest an intermediate situation (which of course is more difficult to model). The protocol data do not suggest that the operator returns to check the control state every time he has met any of the other cognitive needs (though there is always the possibility that this check could be done by eye movement and pattern recognition which are not mentioned verbally). The length of time between each explicit mention of the control state can be measured from the protocol recordings. Figure 5.1.2.2 shows the frequency of different inter-sample intervals, when the control state was unacceptable and acceptable. The data suggest that the inter-sample interval was longer when the control state was acceptable (though the difference between these two frequency distributions is not statistically significant). There are also other studies, e.g. Crossman et al (1974), which do show that sampling rate depends on the size of control error.

Figure 5.1.2.2 : Frequency of different lengths of inter-sample interval, when the control state was unacceptable and acceptable (Bainbridge, 1972, 7.4.3.b).

Several simple mechanisms could be suggested for this interrupting style of updating, which are related to the mechanisms already available, in particular to the control state represented in working storage. Protocol data are probably not able to distinguish between these possibilities.

1. The level of the control state (the main task goal) could act as a time tag which determines how long any processing which is not directly related to the control state will continue, before the control state is checked again.

2. Alternatively, the uncertainty of the working storage representation of the control state might increase with time, as in Figure 5.1.2.3. This certainty could be (unconsciously, see Section 6) checked after each cognitive need has been met. Processing would return to updating if the uncertainty was too high.

Figure 5.1.2.3 ..(5.6K) : Uncertainty plotted against time, with repeated sampling (from Crossman et al, 1974, Figure 3).

There also needs to be a mechanism for returning to the previous point in the processing if the control state is unchanged.

In summary, this section has suggested two general mechanisms for updating the data in working storage.

1. The representation of the general state of the task environment is updated by special purpose 'sequencers' when the OVERALL 'sequencer' decides, on the basis of the control state, to do this type of general state reviewing.

2. The main task goal is checked more frequently. The timing of this updating might be controlled by :
a. the urgency of the control state setting a time control on how long the person thinks about other aspects of the task before returning to check the control state.
b. the representation of the control state decreasing in certainty over time. After meeting each cognitive need the working storage representation of the control state would be checked, and if this internal representation is too uncertain, then an explicit external check of the control state is made.

These mechanisms may be sufficient to account for this operator's behaviour in the electric power control task. It is important to remember however that this task is simple to update. There is only one main control variable, the electric power usage, which develops in a fairly predictable way. Other aspects of the process state evolve slowly, and this evolution can also, within limits, be anticipated. I have not looked into how to expand the mechanisms discussed so far so they would deal with multidimensional sampling. I assume that the extended mechanisms would be related to those for multitasking, which are mentioned in Section 6.

5.1.3. Additional mechanisms

There need to be at least two additional mechanisms to account for other aspects of keeping up with the state of the external world.

In the furnace task, the operator usually only noticed changes in the process state when he actively looked for them, which is accounted for by the mechanisms in Sections 5.1.1 and 5.1.2. However, he did also react, for example he checked what was happening when he heard the teleprinter operating, see Figure 5.1.3. An organism which only noticed what it was interested in or expecting would not live long. As mentioned in Section 4, part 4.5, there needs to be a mechanism by which salient unexpected changes in the environment can override working storage.

Figure 5.1.3 : An example of recovery from an interruption (Bainbridge et al, 1968, Figure 8).

Another interesting feature was first noticed by Beishon (1969), who called it serendipity. This appears to take at least two forms : while looking at or for information needed for one aspect of their task, a person may notice something which is either (a) relevant to, and affects their thinking about, another aspect of the task, or (b) explains a previously noticed but unexplained event. Previous Sections have suggested that working storage maintains an overview which is a summary of data describing the main aspects of the task. These observations of serendipity suggest that the main cognitive needs may also be active when they are not explicitly being thought about, or at least not explicitly being mentioned in a verbal protocol.

5.2. Working storage as a structured construct

The element mechanism shows clearly that what is in working storage is a construct : what is kept in mind temporarily is the result of thinking about the task, not an untransformed representation of input data. Indeed, it has been called working 'storage' to underline this difference from some models of working memory.
This processing element representation also suggests some ways in which working storage is structured :
a. relative to the cognitive needs,
b. by cross-references between cognitive needs.

These cross-references have been symbolised in the Figures in three ways :
1. explicitly, e.g. Figure 2.3.2.
2. by boxes containing single or double arrows (e.g. Figure 3.2.2.1). The arrow is interpreted as pointing to the data needed, which is already available either earlier in the present processing or elsewhere with another major cognitive need.
3. by boxes with a double-line contour (e.g. Figure 2.3.1). This represents a result found by later processing, usually by a 'routine', which is then passed 'up' to the originating cognitive need.

These mechanisms are sufficient to account for working storage in tasks like this furnace power control, which are simple from the structure point of view as the different items to be remembered are independent single variables. Evidence from other tasks shows that the storage mechanism also needs to be able to represent multidimensional constructs and interrelations. The rest of this subsection does not suggest what these mechanisms might be like, it just gives some typical examples of the data which needs to be dealt with by a more complete account.

5.2.1. Multidimensional constructs

A steelworks blast furnace is a good example of a multidimensional process within which the cause-effect relations cannot be directly observed and are not well understood. Hoc (1989) has shown that blast furnace operators build up multidimensional constructs which represent their understanding of what is happening inside the process. These mediate between the variables of the process which can be measured and displayed, and the control decisions which the operators make. Figure 5.2.1 shows some examples of these multidimensional constructs.

Figure 5.2.1 ..(11.8K) : A blast furnace operator's influence graph (Hoc and Samurçay, 1989, Figure 1, English translation from Hoc, 1989, Figure 6).

5.2.2. Relations between variables

Many tasks also involve representing relations between variables, to describe the current task situation.

Figure 5.2.2.1 : Numbers of items about the aircraft they are controlling, remembered by air-traffic controllers (data from Bisseret, personal communication based on Bisseret, 1970).

Bisseret (1970) found that experienced air-traffic controllers could remember an impressive number of facts about aircraft they had just been controlling (Figure 5.2.2.1). He also found that the frequency with which controllers remembered a particular type of fact describing an aircraft (its height, direction, or speed, for example) correlated with the sequence in which these items were thought about while checking for aircraft which were too close together (this sequence was identified by Leplat and Bisseret, 1965). Sperandio (1970) found that the number of items remembered about a particular aircraft depended on how important the aircraft was in the current thinking (see Figure 5.2.2.2). Both these findings might be accounted for by conventional memory models, in which remembering an item depends on rehearsing it. In both studies the number of items remembered is related to the frequency with which an aircraft or item is likely to have been thought about.

Figure 5.2.2.2 : Number of items remembered about air-traffic controllers, relative to the controller's categorisation of the aircraft (from data in Sperandio, 1970).

However, Bisseret also collected some anecdotal evidence about how the controllers remembered the aircraft. He found that the controllers, when asked to draw a map of where the aircraft were, would make remarks like 'I've got one at level 150 which is about to pass beacon RLP and another at level 170 which is about 10 minutes behind so is about here'. This suggests that the controllers remember the aircraft, not in isolation but in terms of the relations between them. The controller has found these relations, such as which aircraft are going in the same direction, their relative height, and the distance between them, as the result of thinking about the task (using the strategy identified by Leplat and Bisseret op cit). The controllers working storage is a construct, as in the steelworks furnace control task. Any representation of the air-traffic controllers working storage (or 'operational' memory, as Bisseret calls it) would have to represent at least these relative positions in space-time.

Chess playing is another task in which the representation in working storage needs to include relations. Chase and Simon (1973) for example found that, when faced with a random layout of pieces on a chess board, both expert players and non-players remembered the same number of items. However, when the pieces were laid out as in a real game, expert players remembered the layout more fully, and remembered it in terms of game-related analytic features, such as attack, defence, and the quality of the positions and therefore likely quality of the players. So the mechanism for working storage needs to be able to contain relational operators.

These points about relations raise a question about the nature of working storage. Section 2, part 2.3 pointed out that the working storage box should not be taken as a location to which items are transferred for storage, but rather as a type of link. The simplest mechanism for working storage would be for it to 'point to' or activate a representation of the relevant item which is already in long-term storage, in the knowledge base (Hebb, 1949). However, such a mechanism would not be able to account for new structures of relations to describe situations which have not previously been experienced. These may be built up in working storage and then become part of the knowledge base. Perhaps two types of mechanism are needed, to account for behaviour in well known and in unfamiliar situations (see Section 7).

5.3. Working storage capacity

Section 3, part 3.2.1 suggested that the 'routines', knowledge bases, and cross references in working storage make a mutually reinforcing redundant structure which supports the working storage. This might enable working storage for task related data to have a larger capacity than is found in laboratory studies of memory for independent context-free items.

It is a pity that there are not many studies of memory capacity for meaningful material in complex tasks. The data in Figure 5.2.2.1 show that experienced air-traffic controllers could recall on average 33 items about the aircraft they were controlling. The evidence on how these items were remembered suggests that these items are remembered in a related way which probably cannot be reduced to a small number of independent chunks. As another piece of anecdotal evidence, air-traffic controllers talk of 'loosing the picture' as a whole, rather than individual items.

Evidence from the furnace operation study (Section 4, part 4.4) suggests that the data found by a main 'routine', for the originating cognitive need, may be available for a longer period of time than just during the 'routine', and may be referred to by other 'routines'. This has been called continuing working storage or the overview, and may be represented in the Figures by a double-line 'head' box. On the other hand, the working storage items within a 'routine' appear to be available to reference only during that 'routine', this type of local storage is temporary. If cross references are only to within a 'routine', or to continuing working storage, this might suggest that a 'routine' is an independent processing module. This is discussed more in Section 8, part 8.2.

In Bainbridge (1972) the furnace operator simulation was used as a basis for making some counts of working storage capacity. These counts distinguished between two types of remembered item:
a. items of 'continuing' working storage. These are the items found by the main 'routines' and then available for reference by other 'routines', as listed in Table 3.2.3.1.
b. items of 'temporary' working storage. These are items which are obtained and cross referenced within a 'routine', but do not appear to be retained for later reference once the 'routine' has been completed, i.e. the originating cognitive need has been met.
(This distinction between continuing and temporary working storage was made in Section 2, part 2.3 and Section 4, part 4.4.)

A count was made, from the simulation, of the number of items of working storage which would be needed while carrying out each 'routine' and 'sequencer'. The counts are given in Table 5.3. These numbers are of course not direct data about human capacities. They have been inferred from the simulation, so their validity depends on the validity of the simulation. For what they are worth, these results suggest that there are about a dozen continuing items (Table 3.2.3.1), but that the number of items actively needed during any one aspect of processing is less than or about seven (Table 5.3).

Routine	repeated	temporary	cross referenced to other	provided for cross reference	total
1	2	3	-	2	6
2	-	?		1
3a	-	2	2	1	4
take action	-	-	2	1	2
4b	-	2-4	2	1	5-7
5a	5?	1?	-	5	6
5b	-	1-3	1-3?	-	1-3
5c	-	2	-	1(3?)	3
6a	-	3	3	1	7
6b	-	1	1	1	2
Sequencer
overall	-	-	2	-	2
unaccept	-	-	3	-	3
step change	-	-	0-3	2	0-3
end time	-	1	2	2	4
omgoing	8-12-14	-	-	8-12-14	8-12-14

Table 5.3 : Number of stored variables needed by each 'routine' and 'sequencer' (Bainbridge, 1972, Table 7.4.1.a).

The data from air-traffic control (e.g. Figure 5.2.2.1) suggest that the number of items that can be recalled about what is happening in a complex dynamic task is much larger than the memory capacity found in laboratory studies of short-term memory for context-free material. However, the number of items of working storage which are actively involved during any one aspect of processing may be less than or about seven (Table 5.3). This might suggest that the ongoing contextual overview which is built up and cross referenced, and the temporary working space actually used during processing, are different mechanisms.

Summary of main points in Section 5

* Updating the data in working storage about the general state of the task, and about the state of the main task goals, is organised by the 'sequencers'.
* In a dynamic task, 'sequencers' and 'routines' may be interrupted to check the state of the main task goals.
* Mechanisms are also needed, for highly salient environmental information to override the overview and its control of processing, and for the way in which people notice information relevant to parts of the task they are not currently thinking about.
* What is in working storage is a construct, the result of processing, not a direct representation of the environment.
* Working storage can contain multidimensional constructs, and relations between variables.
* Working storage for items used within the 'routines' is temporary, these items are only available for cross reference during the 'routine'.
* Working storage for the results of the main 'routines', which is kept in the 'sequencers', is available continuously for reference by other aspects of processing.
* The total number of items that can be recalled about what is happening in a complex dynamic task is much larger than the memory capacity found in laboratory studies of short-term memory for context-free material.
* The number of items of working storage which are actively involved during any
one aspect of processing may be less than or about seven.

©1998 Lisanne Bainbridge