Ch. 15 Acquiring First-order Knowledge About Air Traffic Control Yves Kodratoff and Christel Vrain 발표자 : 권 용 식
Contents. Problem Domain Overall Learning Process Application of ATC
Air Traffic Control(ATC) avoid the crash of an aircraft –if collision is possible, modify the route of the planes AF2470(1) (2) BR1667(1) (2)
Representation of Knowledge and Generalization Trade-off –expressive power –computational complexity Generalization –By using background knowledge BK x y owns(x,y) car(y) driver(x,y) x new_driver(x) driving_license(x) driving_license(John) owns(John, C1) car(C1) new_driver(Bob) owns(Bob,C2) car(C2) driving_license(X) owns(X,C) car(C)
Overall Learning Process determine the concepts to learn acquire background knowledge gather positive and negative examples generalize validate the knowledge Determination of the Concepts Rules: if G(x) then C(x) Positive and Negative Examples Background Knowledge Generalization Tool Validation Concept C(x)
Choose the Language Representation choosing the basic vocabulary choosing the formalism, propositional or predicate, attribute-value representation etc. –ex) (color_eyes, John, Brown) attribute(OBJECT,VALUE) –color_eyes(John, Brown) value(OBJECT, ATTRIBUTE) –brown(John, color_eyes) pred(OBJECT, ATTRIBUTE, VALUE) –physical_descr(John, color_eyes, Brown) attribute(OBJECT, C) value(C) –color_eyes(John, C) brown(C)
Determine the Concepts to Learn The aim of this application is to learn decision rules used by a human controller. Define some actions that can be performed on aircraft to avoid an impending accident –Do nothing: the action is performed on the other aircraft –Change of direction which is a temporary modification of the route –Change of route of the aircraft –... A solution of a conflict is a combination of these actions
Obtain the Examples and Rewriting Them Obtain the “Raw” Examples Rewrite ( (present_work_load PWL1)(steady PWL1) (destination AF2470 DESTINATION1)(west_european DESTINATION1) (destination AF2470 DESTINATION2)(far_away DESTINATION2) (solution AF2470 SOLUTION1)(change_direction SOLUTION1) (right SOLUTION1) (solution BR1667 SOLUTION2))
Rewrite BK as Horn Clauses IF a given plan has a steady trajectory when entering, exiting the sector and inside the sector, THEN it is on cruise x y z t[[traj_enter(x,y) steady(y) traj_in(x,z) steady(z) traj_exit(x,t) steady(t)] fl[flight(x,fl) cruising(fl)]] x y z t[[traj_enter(x,y) steady(y) traj_in(x,z) steady(z) traj_exit(x,t) steady(t)] flight(x,f(x,y,z,t)]] x y z t fl[[traj_enter(x,y) steady(y) traj_in(x,z) steady(z) traj_exit(x,t) steady(t) flight(x,fl)] cruising(fl)]
Generalize and Rewrite G = ( (present_work_load vg3)(steady vg3)(destination vg1, vg21) (destination vg2 vg22)…(solution vg1 vg26)(change_direction vg26) (solution vg2 vg27) Foreseen work load - steady Estimation of the angle of the routes of the aircraft before conflict occurrence - sharp convergence Solution - vg1: change of direction...
Conclusion Transformation of the examples and the expert’s knowledge in first order, and transformation of learning back into a representation are not trivial. The success of the learning process often relies on a proper choice of the formalism Background knowledge is often incomplete, so validation of learning is important. There is still no full automation of ATC.