FACTORIZATION OF LIFEWORLDS

Simon, in Sciences of the Artificial [38], argued that complex systems had to be ``nearly decomposable.'' His model for this was the rooms in a building, whose walls tend to minimize the effects that activity in one room has upon activity in another. Sussman [39], in his analysis of block-stacking tasks, classified several types of ``subgoal interactions'' that result from attempts to break tasks down into subtasks; one hopes that these tasks will be decomposable, but bugs arise when they are not decomposable enough. One assumes that a task is decomposable unless one has reason to believe otherwise. Sussman's research, and the rich tradition of planning research that it helped inaugurate, concerned the difficult problem of constructing plans in the presence of subgoal interactions. Our goal, complementary to theirs, is to analyze the many ways in which tasks really are decomposable, and to derive the broadest range of conditions under which moment-to-moment activity can proceed without extensive analysis of potential interactions.

A non-pathological lifeworld will be structured in ways that limit or prevent interactions among subtasks. Some of these structures might be taxonomized as follows:

• Activity partition. Most lifeworlds separate activities under discrete headings: sewing is a distinct activity from bathing, gathering food is a separate activity from giving birth, and so on. These distinctions provide the basis for reckoning ``different activities'' for the purposes of most of the rest of the partitions. The boundaries among the various activities are often marked through some type of ritual.
• Spatial partition. Different things are often done in different places. Tasks may be confined to the places where their associated tools or materials are stored, or where suitable conditions or lighting or safety obtain. These places may even be close together, as when different recipes are prepared in different sections of countertop space or different kinds of food are kept in different parts of one's plate, with boundary regions perhaps employed to assemble forkfuls of neighboring foods. In general, activities are arranged in space, and decisions made in one place tend to have minimal interaction with decisions made in other places. Of course spatial distance brings no absolute guarantees about functional independence (using up all the resources at one location will prevent them from being carted to another location for another use later on), so these are just general tendencies.
• Material partition. Different activities often involve different materials, so that decisions that affect the materials in one activity do not interact with decisions that affect the materials of the other activity.
• Temporal partition. Different activities often take place at different times, thus limiting the channels through which they might constrain one another. These times might be standardized points during the cycle of the day or week, or their ordering might be constrained by some kind of precondition that the first activity produces and successive ones depend upon.
• Role partition. Simon pointed out that division of labor eases cognitive burdens. It does this in part by supplying individuals with separate spheres in which to conduct their respective activities.
• Background maintenance. Many activities have background conditions that are maintained without reference to specific goals. For example, one maintains stocks of supplies in the pantry, puts things back where they belong, and so forth. Hammond, Converse, and Grass [16] call some of these ``stabilization.'' (See Section 5.) What these practices stabilize are the relationships between an agent and the materials used in its customary activities. They tend to ensure, for example, that one will encounter one's hammer or the currently opened box of corn flakes in definite sorts of recurring situations. They thus reduce the complexity of life, and the variety of different hassles that arise, by encouraging the rise of routine patterns and cycles of activity rather than a constant stream of unique puzzles.
• Attributes of tools. Numerous properties of tools limit the interactions among separate decisions. Virtually all tools are resettable, meaning that regardless of what one has been doing with them, they can be restored to some normal state within which their full range of functionalities is accessible. (This of course assumes that one has only been using the tools in the customary ways and has not been breaking them.) Thus the properties of the tool do not place any ordering constraints on the activities that use it. Likewise, most tools are not committed to tasks over long periods. Once you have turned a screw with a screwdriver, for example, the screwdriver does not stay ``stuck'' to that screw for any long period. Thus it is not necessary to schedule the use of a screwdriver unless several people wish to use it at once. Exceptions to this general rule include bowls (whose ingredients must often sit waiting for future actions or conditions, and which cannot contain anything else in the meantime), stove burners (which sometimes must remain committed to heating particular dishes until they have reached certain states and not before), and clamps (which must remain fastened until the glue has dried or the sawing operations have been completed).
• Supplies of tools. These latter tools raise the spectre of generalized scheduling problems and the potential for deadlock among multiple activities, and such problems do in fact sometimes arise when cooking for more people than the number to which a given kitchen is adapted. Most of the time, though, one solves such problems not through scheduling but simply through having enough of the tools that must remain committed to particular purposes over a period of time. Lansky and Fogelsong [22] modeled the effects on search spaces of limited interactions between different cooks using overlapping sets of tools.
• Warning signs. When things go wrong, unpleasant subgoal interactions can ensue. To avoid such difficulties, an individual, community, or species keeps track of warning signs and cultivates the capacity to notice them; these warning signs include supplies running low and funny smells. This is often done on a primitive associative level, as when rats stay away from smells that were associated with stuff that made them sick or people develop phobias about things that were present when they suffered traumas. Communities often arrange for certain warning signs to become obtrusive, as when kettles whistle or natural gas is mixed with another gas that has a distinctive smell.
• Simple impossibility. Sometimes things are just impossible, and obviously so, so that it is not necessary to invest great effort in deciding not to do them.
• Monotonicity. Many actions or changes of state are irreversible. Irreversible changes cause decisions to interact when certain things must be done before the change takes place. But it also provides a structure for the decision process: the lifeworld needs to make it evident what must be done before a given irreversible change occurs.
• Flow paths. Often a lifeworld will be arranged so that particular materials (parts on an assembly line, paperwork in an organization, food on its way from refrigerator to stove to table) follow definite paths. These paths then provide a great deal of structure for decision-making. By inspecting various points along a path, for example, one can see what needs to be done next. Or by determining where an object is, one can determine what must be done to it and where it must be taken afterward. Some of these paths are consciously mapped out and others are emergent properties of a set of customs.
• Cycles. Likewise, many lifeworlds involve stable cycles of activities, perhaps with some of the cycles nested inside of others. The resulting rhythms are often expressed in recurring combinations of materials, decisions, spatial arrangements, warning signs, and so on.
• Externalized state. To computer people, ``state'' (used as a mass noun) means discernible differences in things that can be modified voluntarily, and that can be interpreted as functionally significant in some way. Early AI did not treat internal state (memory) and external state (functionally significant mutable states of the world) as importantly different, and it is often analytically convenient to treat them in a uniform fashion. It is often advantageous to record state in the world, whether in the relative locations of things and the persistent states (in the count noun sense) that they are left in [7]. For example, one need not remember whether the eggs have been broken if that fact is readily perceptible, if one's attention will be drawn to it on a suitable occasion, and if one understands its significance for the task. Likewise, one can save a great deal of memory by retrieving all of the ingredients for an evening's recipes from the cupboards and placing them in a customary place on the shelf.

Lifeworlds, then, have a great deal of structure that permits decisions to be made independently of one another. The point is not that real lifeworlds permit anyone to live in a 100% ``reactive'' mode, without performing any significant computation, or even that this would be desirable. The point, rather, is that the nontrivial cognition that people do perform takes place against a very considerable background of familiar and generally reliable dynamic structure.

The factorability of lifeworlds helps particularly in understanding the activities of an agent with a body. A great deal of focusing is inherent in embodiment. If you can only look in one place at a time, or handle only one tool at a time, your activities will necessarily be serial. Your attention will have a certain degree of hysteresis: having gotten to work on one countertop or using one particular tool, for example, the most natural step is to carry on with that same task. It is crucial, therefore, that different tasks be relatively separate in their consequences, that the lifeworld provide clues when a change of task is necessary, and that other functionally significant conditions can generally be detected using general-purpose forms of vigilance such as occasionally looking around. Of course, certain kinds of activities are more complex than this, and they require special-purpose strategies that go beyond simple heuristic policies such as ``find something that needs doing and do it.'' The point is that these more complex activities with many interacting components are rare, that they are generally conducted in specially designed or adapted lifeworlds, and that most lifeworlds are structured to minimize the difficulty of tasks rather than to increase it.

These various phenomena together formed the motivation for the concept of indexical-functional or deictic representation [4, 3]. Embodied agents are focused on one activity and one set of objects at a time; many of these objects are specifically adapted for that activity; their relevant states are generally readily perceptible; objects which are not perceptibly different are generally interchangeable; and stabilization practices help ensure that these objects are encountered in standardized ways. It thus makes sense, for most purposes, to represent objects in generic ways through one's relationships to them. The flashlight I keep in the car is the-flashlight-I-keep-in-the-car and not FLASHLIGHT-13. I maintain a stable relationship to this flashlight by keeping it in a standard place, putting it back there when I am done with it, using it only for its intended purposes, keeping its batteries fresh, and so on. Its presence in the environment ensures that I have ready access to light when my car breaks down at night, and therefore that I need not separately plan for that contingency each time I drive. The conventional structures of my own activity maintain the flashlight's presence as a ``loop invariant.'' Both the presence of the flashlight and the activities that ensure it are structures of my lifeworld.