Intelligent Systems And Their Societies Walter Fritz



ICARUS was reported on by Pat Langley, Kathleen McKusick, John Allen, Wayne Iba, and Kevin Thompson of the AI Research Branch of the NASA Ames Research Center. It is described in the SIGART journal of August 1991.

This is an intelligent system that was to be used for controlling a physical robot, allowing it to move around by itself and to manipulate objects, possibly on another planet. ICARUS is composed of asynchronous parts. That means, parts that all work separately and interchange information through a type of "blackboard". These parts are a perception system, a planning system, an execution system, and a memory system. Each reads information from the central "blackboard", works on this information, and, as a result, normally posts information back to the "blackboard". The exception to this is the execution system which may execute a primitive response instead (e.g. run a certain motor for a given time).

The perception system receives sensory information, uses a mechanism that centers attention, and posts the present situation on the blackboard expressed as a qualitative description of objects and events (we would say expressed as concepts).

ICARUS's planning system is bi-directional recursive and works with response rules (they call them "operators") that try to eliminate the difference between the present and objective situation. That is, it works forward from the present situation trying to reach its current objective situation and also backwards from the objective situation trying to reach the present situation. If the two newly generated situations are not identical, they become the "present situation" and "objective situation" for the next iteration. When a match is finally found, the cumulative plan is posted on the "blackboard".

Beyond this basic mechanism, the planning system also adds an additional function that increases its efficiency. The planning system may interrupt itself if it finds that a part of the present (and changing) situation represents a high priority problem for which a solution already exists in memory. This may also happen if the human operator posts a new, high priority objective.

The execution system takes the current plan from the blackboard and executes the response sequence. If, during execution, an unforeseen intermediate situation is created and recognized, the execution system stops and waits for a new or updated plan from the planning system.

The memory system gains knowledge from experiences. It creates composite and abstract concepts whose parts can also be concepts. For instance, it creates composite concepts from several concepts (that are objects) and their relationship. Together, all its concepts form a probabilistic hierarchy of concepts.

This means that there is a probability attached to each part of a concept that indicates how probable it is that this part belongs to the concept. At the lowest level it bases all its concepts on something that it observes in its environment. It also creates new response rules based on the present situation, the final situation, and the relationship(s) between them.

Curiously, ICARUS makes no great distinction between response rules and concepts; it represents both in the same format. (ICARUS is the only program that I know of, where this is the case.) Both are managed as relationships (as patterns) even though concepts are a relationship over space and rules are a relationship over time. (In the human brain there is also no differentiation, both are patterns of neuron connections. However, it is in the human mind that this differentiation occurs.)


For continuous reading, like a book - continuehere
Jump to the top of the ICARUS document /Artificial IS document /e-book Contents.

Last Edited 2 Mayo 2013 / Walter Fritz
Copyright © New Horizons Press