![[Home]](/ailab/people/andel/pics/AIWiki.png)
Langley et al 1991 defines the AgentArchitecture, and Shapiro et al 2001 or Daniel Shapiro ‘s thesis defines the AgentLanguage? for specifying the behavior of an Icarus.
The AgentLanguage? is a reactive hierarchical behavior description. It is related to TeleoReactiveTree s and UniversalPlan?s.
For an example Icarus plan, see IcarusDrivingExample.
“An Icarus plan contains up to three elements: an objective, a set of requirements (or preconditions), and a set of alternate means (or methods for achieving objectives)… Each of these can be instantiated by further Icarus plans, creating a logical hierarchy that terminates with calls to primitive actions or sensors. Icarus evaluates these fields… beginning with the objective field. If the objective is already true in the world, evaluation succeeds and nothing further needs to be done. If the objective is false, the interpreter examines the requirements field to determine if the preconditions for action have been met. If the objective is false and the requirements are true, evaluation progresses to the means field, which contains alternate methods (primitive actions or subplans) for accomplishing the objective.”
“As the interpreter encounters sensor tests it selectively descends the calling tree, ultimately locating one or more relevant actions.
As the calling stack unwinds, Icarus returns the best action found in each subplan. The action returned from the top-level plan is passed to an external execution system, which applies it in the world.
Thus, the purpose of an Icarus program is to find action.”
A paper about using Icarus for BehavioralCloning is Ichise et al 2002. The procedure they give successfully learned the above behavior from training examples.
Ichise, R., Shapiro, D. G., & Langley, P. (2002). Learning hierarchical skills from observation. Unpublished manuscript, Institute for the Study of Learning and Expertise, Palo Alto, CA.
P. Langley, K. McKusick?, and J. Allen. A design for the ICARUS architecture. In ACM? SIGART? Bulletin, pages 104--109. ACM? publications, 1991.
Shapiro, D., Langley, P., & Shachter, R. (2001). Using background knowledge to speed reinforcement learning in physical agents. Proceedings of the Fifth International Conference on Autonomous Agents (pp. 254-261). Montreal: ACM? Press.