We believe that context impacts behavior in at least five different ways. First, recognizing the context it is in allows an intelligent agent to make predictions about the situation that biases situation assessment. This has certainly been shown in experiments with humans [see, e.g.,]glass:holyoak. Contextual information provides top-down predictions that allows recognition of important objects and attributes based on partial knowledge about them, and it can serve as a source of hypotheses about objects, states, or events not yet seen. For example, an AUV controller may realize that it is near a constriction between two bodies of water, such as often occurs between an estuary and the ocean; in such a context, strong currents are expected. This prediction would allow the AUV to quickly determine what is happening when it is having difficulty maintaining its heading, allowing it to rule out equipment malfunction.
Second, context modulates behavior by setting an agent's behavioral parameters and suggesting which goals (or actions) should be activated or deactivated. For example, when entering a library, a person automatically lowers his or her voice; when entering a classroom, a student automatically has the goal of finding a seat, while the instructor has the goal of moving to the front of the room. Context-specific modulation of behavior is also important for artificial agents. For example, when an AUV enters a harbor, it should realize without a lot of reasoning effort that in this context, it needs to tighten its ``depth envelope''-it should stay away from the surface, since there is likely to be surface traffic there, and it should also stay away from the bottom, which is likely to be shallow.
Third, context helps an agent focus its attention. At any given time, an agent will often have more goals than it has cognitive and other resources to achieve; consequently, it needs some way of determining which of its goals is (or are) the best to work on-but this differs depending on the context. For example, when driving home from work, it is perfectly appropriate for an ambulance driver to pick up his or her laundry; however, when taking a patient to the hospital, the ambulance driver should not even consider this goal, though it may still be present. Similarly, an AUV may have the recurrent goal of letting its base know where it is; if the AUV is operating in the open ocean near the surface, this is a reasonable goal to pursue from time to time; however, if the AUV is operating near the bottom of the ocean or under ice, then this goal is no longer a reasonable one to pursue, since it requires surfacing.
Fourth, context influences an agent's choice of actions for a particular goal. The appropriateness of an action cannot be assessed apart from the context in which it is to occur. For example, the way one chooses to get a drink of water differs depending on whether one is at home, in a restaurant, or at a friend's. Similarly, an AUV's goal of communicating with another AUV will be satisfied differently in different situations: in clear water over relatively short distances, a laser may be used for high-bandwidth communication; in murky water, however, or over large distances, an acoustic modem may have to be used instead.
Fifth, context determines how an agent should handle unanticipated events. ``Unanticipated event'' here does not mean one that is novel or completely unexpected, but rather one that cannot be completely predicted ahead of time. A driver knows that one of the things that can happen while driving is that a pedestrian can step in front of the car; however, when a particular pedestrian does so at a particular time, it is an unanticipated event. Events have different meaning in different situations. Seeing a red flashing light through your office window will affect you differently than seeing one in your rearview mirror while driving, both in terms of your heartrate and your reactions. Such differential responses to events based on context have value; it would be unwise to always ignore a nearby gunshot, yet if one reacted the same way when at a firing range as when walking down the street, it would be exhausting (as well as annoying for others). For an autonomous agent such as an AUV, for which reactions must often be fast yet appropriate, it is crucial that some means exist for context to automatically condition the agent's responses to events. For example, when power is about to fail, an AUV will have very little time to choose its course of action, yet what is appropriate differs based on the situation. If the AUV is in the open ocean, the best response may be to surface and signal for help; in a harbor, where surfacing would risk getting run over, landing and releasing a buoy may be best; and when tethered to a support vessel during launch or recovery, the best response may be to do nothing or to send a signal via the tether to the human operators.