Previous: Related work Up: Related work Next: Implemented Architectures
We believe that behavior-based AI has adopted the right treatment of every day behavior for agents that function in the world. However, this has been done at the expense of ignoring cognitive processing such as planning and reasoning. Clearly, what is needed is an approach that allows for both. We believe that our architecture meets this need. As in behavior-based AI, GLAIR gains validity from its being grounded in its interaction with the environment, while it benefits from a knowledge level that, independent of reacting to a changing environment, performs reasoning and planning.
The Model Human Processor (MHP) is a cognitive model [Card et al. 1983] that suggests the three components of perception, cognition, and motor. Cognition consists of working memory, long-term memory, and the cognitive processor. Perception is a hierarchy of sensory processing. Motor executes the actions in the working memory. This is a traditional symbol-system decomposition of human information processing. This type of decomposition has shown only limited success in building physical systems. Despite this, systems like SOAR adhere to this model. In our architecture, we purposively avoid this kind of top-down problem decomposition by allowing independent control mechanisms at different levels to take control of the agent's behavior, and pre-empt higher level control while doing so. It may be necessary to allow higher level mechanisms to selectively inhibit lower-level ones as well, but we have found no good reason to do so yet.
A situated agent, at any moment, attends to only a handful of entities and
relationships in its immediate surroundings. In this type of setting, the
agent often does not care to uniquely identify objects. It is sufficient
to know the current relationship of the relevant objects to the agent, and
what roles the objects play in the agent's activities. Agre and Chapman in
[Agre \& Chapman 1987] proposed indexical-functional representations (which
[Agre 1988] refers to as deictic representations) to be
the more natural way agents refer to objects in common everyday
environments. They called entities and relationships of interest
entities and aspects, respectively. With respect to its current
activities, the agent needs only to focus on representing those entities
and relationships. Although the objects in the environment come and go,
the representations of entities and relationships remains the same. For
example, the-cup-that-I-am-holding is an
indexical-functional notation that abstracts the essentials of what the
agent needs to know in its interaction. These representations serve to
limit the scope of focus on entities. For example, if the agent wants to
pick up a cup, it does not need to know who owns the cup or how
much coffee the cup can hold; only the relevant attributes of the cup
apply. We believe that systems endowed with
general KRR abilities can and should generate deictic representations to
create and maintain a focus on entities in the world, but we have not yet
designed an implementation strategy.