Knowledge in ICARUS: LABYRINTH

LABYRINTH stores knowledge in a hierarchical tree using probabilistic concepts to define the nodes of the tree. Layers of the tree correspond to levels of abstraction of the concept and nodes at a particular level of the tree store the probability that the role represented by the node is contained within a specific instance of the abstract concept (given that the object is a member of the parent concept). In other words, an abstract concept is denoted by a high-level node while a specific event is denoted by a terminal node. The structure that one passes through between an abstract concept and a terminal node indicates the features of the object and, through the probabilities assigned to the roles, how representative the object is of the class in general.

LABYRINTH stores experiences as objects, states, and plans, where each component of an experience is sorted throughout memory to determine the best child node to incorporate the experience, or to store novel experiences as new nodes. At each storing, LABYRINTH classifies the composite based on classifications of each component, as well as relations between components. All information is stored hierarchically, bottom-up from components of an object (for example), upward to the entire concept. Furthermore, statistics about objects, plans, and states are stored to determine the better problem-solving methods or better plans.

There is a distinction between long-term memory and working memory. Working memory is the portion of the concept hierarchy through which experiences have recently passed. Long-term memory is everything else. This style of organization--both the hierarchical approach and the short-term/long-term distinction--allow for fast access to knowledge in hopes of decreasing the reaction time of the system. This methodology is also associative in nature as well.

To return, press HOME