Interruptions in the Workplace - DRAFT
David Kirsh, UCSD
Viewed abstractly we may understand interruption as a break in our organization of activity arising from a change of (task) environment. Whether an interruption is caused by having to pick up a telephone and begin a conversation, by the need to honor social obligations and acknowledge visitors who arrive at our office door, or whether it is generated internally by our own decision to shift what we are doing -- as when we switch from one computer program to another to pursue a sub task such as compose a graphic for the paper we are working on, or even just stop the writing process in order to correct words or fix formatting -- we are shifting our attention from one phase of activity, with its attendant resources to another. This shift of focus involves a context switch. The result is that we must leave our current task environment -- the task we are highly situated in -- and enter a new one. Just how disruptive that context switch is depends on the infrastructure in each environment and how we leave and enter.
In the psychological literature these last few years there has been an increase in the number of articles published on context switching. For the most part these studies have focused on the need to switch between two formal tasks, such as we find in split attention studies. Unlike traditional split attention tasks, where the tasks which attention is divided between are quite different, in the psychological literature, the tasks are different but the stimuli are the same. For instance, one set of studies examined how long it takes someone to switch between naming by word and naming by color in Stroop like contexts. For instance, given a set of three words each colored using a different font color -- RED, GREEN, ORANGE -- read the color word, then tell us the font color of the word. Though potentially helpful, in my opinion, these careful psychological studies are not yet sufficiently ecologically natural and molar to suppose that they will teach us much about interruption in the workplace.
I can be more explicit. The bottom line from a psychological perspective is that ultimately the factors that determine how intrusive or disruptive a context switch will be depends on three things:
When people switch context they have to reconfigure the procedures that are active in Working memory, they have to redefine the cues that are meaningful in the context, they have to figure out where they are in the task, and so forth. Just how these elements of task switching play out in the psychological domain is the job of psychologists to determine. But though one would expect classical cognitive psychologists to have a great deal to say about these topics there is not nearly as much known here as one would expect.
Psychologists have studied human and animal memory extensively. But typically their focus has been on how the memory system works --its capacities, its strategies, its processes. In the long run this will be helpful, but often research is at too micro a level to help us understand interruption in the workplace, because we would like to know how memory works in a highly situated way. That is, how memory is put to use in ongoing naturally occurring activities, not how it is used in highly artificial tasks that are not ongoing parts of subjects' everyday lives. Despite the need for careful and highly controlled experimentation, we need to connect our laboratory experiments with ecologically natural contexts of the sort we find knowledge workers in everyday.
One way of appreciating how important ecological reality is here is to ask what is involved in recovering from an interruption. At first it might seem that the problem is one of recovering the `task set' one was in before the interruption. Some of the scaffolding for performing the task is in the environment and some is in the head of the agent. The question then is to determine how the subject recovers task set by relying on external cues and how he or she recovers by relying on memory. There is much more to be said here about the nature of projecting structure, about creating systems of salience, about inventing strategies, and so on. But in fact, the problem is more complex. For in real settings it is typically the case that at any moment there are many tasks which one might be choose to perform and the trick is to decide which to start or restart. This makes problem choosing or task selection to be a significant element in studies of interruption.
One way into this problem for ecological studies is to create a table as follows:
In this table we collect as many to do's as we can, and assign them an importance value/rank, a hardness value, and we try to determine which of the necessary resources are at hand. We then observe how people actually behave after they are interrupted. Do people turn to tasks that are ready to be done whereas they are more reluctant to turn to tasks that require preparation? This represents an entire line of observational and computational research.
Another line of research concerns the impact of interruption as a function of where in a task you are. If people are interrupted when they are just starting out in a task the disruptive nature may be greatest, sicne they are setting up the environment. Moreover, there is a lot that people can do to make interruptions less intrusive, or at least, there is a lot they can do to make the consequences of an interruption less negative. Some of what can be done is to stabilize one's memory at the moment of interruption by, for instance, noticing where in the task one is, or what sub goal one has reliably achieved, or where one should pick up the task when it is time to return. This stabilization may involve consciously shutting down momentary processes of problem solving, that involve rapidly changing working memory states, or it may involve consciously noting elements of the environment that can serve as cues to bring one back into the task. Such notions of what can be done are speculative, however, until we have a better theoretical understanding of the ecology of interruption; of the way control of workflow is distributed over internal states, resources, and other people in our work environments.
In this note I will briefly discuss how I typically think about how state is held in humans, in their work environment and in their distribution of effort with their colleagues. I do not mean this to be scholarly or complete, though I believe that it does help to set the stage to how to think in an disciplined manner about interruptions in ecologically natural contexts. This position paper will set the view for a brief discussion of the literature, such as it is.
There are various memory systems simultaneously involved in keeping track of where we are in a task. Our visual system has memory for the spatial arrangement of the items in the visual field. We have a longer term memory for the spatial layout of the environment we are in. Our auditory system builds an auditory map of the auditory scene and this feeds into our memory for spatial layout and also of the characteristic ambient sounds normally found there. We rely on our prospective memory to keep track of goals and actions we intend to take in our task, and we also rely on autobiographical memory to track our episodic involvement in our environment - our memory for the places, times and occasions when we were involved in concrete activities in getting to the spot we are currently. And finally we rely on our semantic or non-episodic memory for programs, ideas and concepts about our tasks that we do not tag to particular episodic experiences.
Many of these memory states are highly situated. Thus our memory for spatial layout is not necessarily map-like, it may be procedural and landmark based. We know where to expect to find the knife we put down after chopping, but our recall for that location is relative to the chopping board we were using it with just before. The same can be said for our episodic recollection. Our recall of how we reached the task state we are in, will naturally be biased by the first person orientation we have to the task we are performing. This means that although we may understand the states we created in an abstract way, our initial encoding of the states contained our spatial perspective, and our context sensitive understanding of the state at the time. I recall chopping the tomatoes and crumbling the feta cheese, but in order to remember whether I noticed whether we were out of olive oil at that time I have to recreate my involvement in the task from the perspective I had at the time. The same also applies to prospective memory. If I am making a stew and salad I may intend to stir the stew as soon as I have finished the salad. But if I am interrupted, I may forget that I had intended to stir the stew after completing the salad because I may actually have encoded my prospective intention to cue off of a specific event in salad completion, such as tossing the salad or pouring on the dressing. Often, what I have stored in memory is not an absolute intention but a situated one.
In a related way our familiarity with ambient sound is also situated. If we walk into a setting which we know well, we expect to hear certain sounds. If these sounds are absent or altered we very soon notice that something is unusual. But the sounds we expect to hear have to do with the activity we see, as well as the path we take. Our memory for normal sound is situated in a network of other elements to do with our expectations of normal activity, the way we move through a room and so on.
Finally, there is working memory - the short term memory required by the reasoning, planning, and behavior management processes we are currently involved in. Since such memory lasts at most a few seconds it is situated in an even more powerful way. It is meaningful only with respect to the cognitive and behavioral processes occurring at the time.
When an interruption occurs, all our limited resource systems are deeply effected. Using computer metaphors, information is swapped out of each limited memory system and if possible a trace is placed in long term memory. The frame of action we depend on to know what we are currently doing, is `paged' out and a new one created or paged in. Thus, if the sound of a timer alerts us to remove the pie in the oven and so interrupts our current sub-task of cutting an onion, we must swap out the various memory states associated with the knife, onion and cutting board, and swap in the appropriate activity frame associated with the buzzer and removing a hot pie from the oven. When the pie is out we need to remember to return to the job of cutting, and continue the process where we left off.
There is much more to be said here, but these few paragraphs should show the orientation of my thought on the topic.
Information about the current state of a task is also encoded and held in the environment in a variety of ways. Consider cooking again. Whether cooks realize it or not, their activity can often be broken into three phases:
1. a preparation phase, where they bring out the ingredients and then chop and semi-process them into the form they must be in in order to serve as input for the main phase of cooking,
2. a cooking or transformation phase where the semi-processed ingredients are mixed together and heated until done, and
3. a wrap up or stabilizing phase, where they secure the completed product in a safe condition and put it in a safe place before laying it out on a plate to be taken to the dinner table.
In the frenzied activity of real kitchens these phases are not always completely distinct, nor are they easy to isolate since different dishes are being prepared simultaneously and so counter space is usually at a premium. Nonetheless, if an expert chef were to step into a kitchen at the end of any one of these phases it is likely that he or she would have a good idea of what is going on. From a complete layout of ingredients he likely could guess the dish to be prepared. From the preparations in pan or oven, and the time left on the timers nearby, he could guess the likely state of the dish, and from the completed items lying on kitchenware awaiting proper arrangement on finer dinnerware, he could guess their likely final deployment. The matter is slightly more complex in large kitchens as there a chef must be able to recognize the different layers of state in the superposition of multiple meals. But with experience this type of recognition ability is not out of reach. Of course, a novice cook would have no such ideas about what was going on. To be able to read off the state of a process from the arrangement of items in the environment depends on knowing enough of the characteristic states of that process to crack its code. So a novice is like a person who cannot read a description in a foreign language. The information is there but not in a form that he can use.
Cooking is an activity where most of the important structuring events occur in the environment. But as with any process in which there is a controller, there are moments when most of the state is externalized and moments when most of the state is internalized. For instance, in laying out milk, flour, eggs, sugar, and vanilla any baker knows there is a good chance that a cake is about to be made. But within the category of cakes there is a range of recipes. If the precise measurements have not been completed, no amount of baking knowledge can substitute for what the chef has in mind. The state of the task is distributed between chef and environment. Similarly, if there is a page in a recipe book open, much of the state of the cooking world is transparent, but after the book is closed and before the ingredients are taken out, all the crucial state is in the agent's head.
Aside from the `natural' meaning which an environmental state can hold, there are many other ways an environment can hold state information. In a lovely study of a hamburger restaurant, Alex Kirlik observed how short order cooks laid out hamburger paddies on the griddle. The hamburger cook's problem is straightforward. Orders arrive at random intervals for hamburgers that are to be done either rare, medium or well. Since there may be as many as ten or fifteen hamburger paddies on the heat at once, it is a challenge to keep track of how well cooked each is at the moment, and how well cooked each is supposed to be, when it is completely cooked. The solution they have struck upon is ingenious. Partition the griddle into three spaces. The top third represents the well done space. The middle third represents the medium space. And the bottom third represents the rare space. Paddies all are started from the right side of the griddle and are shifted across at a constant rate across the hot surface. The one thing that changes is how far each is moved. Rare burgers are moved in big steps, medium burgers in medium steps and well done burgers in small steps. In this way, there is simple update routine that ensures that the state of the environment can be read off. Anyone who knows the routine can read off how done each burger is, relative to its final state.
Descriptive complexity of the environment is a construct that has yet to be developed but which may have a powerful role to play in discussions of environments. The simple notion is an adaptation of Kolmogorov complexity. Namely, the complexity of an state or layout or structure, is given by the length of the minimal program sufficient to generate that structure. This has to be adapted to the human case since not all programs are meaningfully run by humans, and of course, the complexity if a program also depends on how many cases or structures an agent has in memory. I mention this here only to mark it as a topic to cover in a different essay.
Another topic that must be covered is the importance of stabilizing an environment as you leave it.
I mention this here only to mark it.
Interruption occurs when we leave one environment, or one state of an environment and we are forced to discontinuously jump to another state or enter another environment. The causes of these jumps may be internal - as when I suddenly remember I have to be downtown in a few minutes - or they may be external, as when the doorbell rings or a colleague enters my office with a question. In all cases the challenge is to minimize the disruptiveness of the interruption so that we can pick up the new task quickly (that is, adapt to the new environment quickly) or so we can return to our original environment in a fit state to pick up where we left off.
Designers have two key questions:
1. how can they provide resources and scaffolding to make recovering our memory state as simple and as complete as possible?
2. How can the environment we are leaving be stabilized so that we do not lose any progress we may have made already? Since some interruptions disrupt a process needing human management, most programs fall back to the previous stable state the task was in when the human must leave the program. How can we do better?
There are a few things to be said about each question.
Memory. Two principles of recall that have the most direct application to environment design are the Principle of encoding specificity and the Principle of process specificity. Both are similar in spirit. In a nutshell they state that if you wish to recall an event the easiest way to increase your probability is by recreating as much of the external context of encoding as possible - this is the Principle of encoding specificity - and also by recreating as much as possible of the internal context of encoding - -this is the principle of process specificity. That means that if I have been interrupted in the middle of preparing a cous cous royale, I have the best chance of recalling exactly what I was in the middle of doing at the moment of interruption if you can recreate the world as it was at the moment of my interruption. Of course, it might be even more helpful to recreate the few moments before, so I have a larger specific context to recall. But the potency of the technique depends on having the person relive as much of the previous conditions as possible. Since a component of those previous conditions would be what other processes were going on in the person at the same time, this technique is even more potent when we actually do some of the behavioral and cognitive things we had been doing at the time of the interruption. In that way we recreate for ourselves as much of the internal conditions as well as the external ones as possible.
Again this is just the beginning of large topic, but it is worth laying down the first words.
Interruption is huge topic that until recently has not been appreciated as a major issue for psychologists and cognitive scientists. The cognitive scientist brings to the equation a concern for ecological reality that moves the primary questions away from shift costs, reconfiguration cost, residuals costs, to questions about the attractiveness of tasks in a matrix of activities and needs that each person carries with them.