Post on 02-Apr-2018
D.A. Spencer
Applying Artificial Intelligence Techniques'to Air Traffic Control Automation
We have developed a computer program that automates rudimentary air traffic control(i\TC) planning and decision-making functions. The ability to plan. make decisions,and act on them makes this experimental program qualitatively different from the moreclerical ATe software currently in use. Encouraging results were obtained from testsinvolving simple scenarios used to train air traffic controllers.
The primary responsibility of air traffic control (ATC) is the prevention ofaircraft collisions.An important secondary responsibility is toexpedite traffic. These and other duties are allcurrently performed by human air traffic controllers. Increasing nrnnbers offlights, however,are straining the system-a trend that is expected to continue. Consequently, the FederalAviation Administration (FAA) has turned tocomputer-based aids to assist controllers inperforming certain ATC functions.
The current generation ofsuch aids came online in the early 1970s. The aids can be characterized as clerical in nature; Le., they helphuman controllers to envision current and future traffit:; situations, and they provide information needed for directing aircraft. With a fewverylimited exceptions. however, the current automation system does not recommend actions tocontrollers.
Planned hardware and software updates [1]should improve these aids and lead to furtherincreases in controller productivity. Nonetheless, within the foreseeable future the possibilities for improving the automation of purelyclerical functions will be exhausted. Thus further productivity increases will reqUire aids forautomating the planning and decision-makingprocesses. To that end. the FAA has embarkedon the Automated En Route ATC (AERA) andTerminal ATC Automation (TATCA) projects[2-4]. The FAA's goal is to develop planning anddecision-making aids for directing aircraft thatare flying both en route and within an airportterminal's airspace. Under the third phase of
The Llncoln Laboratory Journal. Volume 2. Number 3 (1989)
AERA, the FAAhas also considered the possibility of totally automating certain ATe functionsor providing totally automated ATe in somesectors of airspace. Unfortunately, very littleprevious research is applicable to the deSign ofATC planning and decision aids, particularly tothose aids which must operate autonomously.
At Lincoln Laboratory, we have developed anartificial intelligence (AI) program that automates rudimentaryATC planning and decisionmaking functions. We tested this experimentalprogram by applying it to two training scenariosthat were used for the on-the-job training of airtraffic controllers at an en route ATC facility.Without manual assistance, the AI programhandled both scenarios successfully, and wasable to keep up with the first (and easier) scenario inreal time. Itis important to note that ourintent was to develop a program with a novicelevel of proficiency sufficient to handle the twotraining problems. The available resources didnot allow the development of a robust expertcontroller. However, the AI system was structured so that, given further development, itcould evolve into a true expert controller.
Ourresearch differs from mostpreviousworkin that we considered an en route controller'stotal job. not just particular subfunctions. (Seethe box "A Brief History ofArtificial IntelligenceApproaches to ATC.") The AI program that wedeveloped makes plans and acts on them.This capability distinguishes the program fromthe current generation of operational ATCautomation tools, which, as stated earlier,simply perform clerical tasks that aid the
537
Spencer - Applying ArtifLCial Intelligence Techniquesto Air Traffic Control Automation
A Brief History of Artificial IntelligenceApproaches to ATe
R.B. Wesson [1] was the first toapply artificial intelligence (AI)techniques to ATC. His work influenced much later research. including our work at Lincoln Laboratory. Wesson attempted toautomate the complete controller's function rather than focusing on one aspect of the controller's job. As a means of testing hissystem, he applied it to SimulatedATC problems that are used totrain controllers at en route sites.
Air Route Traffic Control Center (ARfCC) training problemsrequire controllers to handle asector of airspace under a simulated traffic load for approximately one hour. The sector is areal-life sectorwithin the region ofthe trainee's ATC facility. Although the traffic is simulated,the training problems include realistic situations for that sector.The training program for new controllers involves progressingthrough a sequence ofthese problems. The initial problems haverelatively light traffic; the laterones create a workload heavierthan what is expected underthe most demanding real-worldconditions.
Wesson was able to implementenough of the controller's functions to handle successfully thebasic aspects of several of theproblems. Our work attempts toreproduce Wesson's results (withdifferent scenarios and a somewhat different planning technique) and tries to provide a basisfor developing a completely automated air traffic controller.
Wesson later worked on aRand Corporation team that usedATC to study distributed expertsystems for planning and control[2,3). Since the team focused onalternatives to the current geo-
graphically distributed controlsystem, much of the research isnot applicable to our work because practical considerationsconstrain the existing ATC system to evolve gradually from itscurrent state.
The Rand team also analyzedthe proposed Automated EnRoute ATC (AERA) system [4] forthe Federal Aviation Administration (FAA). From that work, apaper by Wesson [5) resulted.Using both analysis and dramatization, the paper provides a persuasive vision of how the currentATC system could evolve into thelargely automated system that isplanned as the last phase ofAERA.
FollowingWesson's work. S.E.Cross (6) applied techniques fromthe qualitative-physics area ofAIresearch to represent the understanding a controller might haveof the constraints that aerodynamics places on ATC actions.Another interesting aspect ofCross's thesis is a technique fordecomposing a multiple-conflictscenario into a fundamental orminimal set of conflicts that needto be resolved. Like Wesson.Cross also applied his system tosolVing ATC training-scenarioproblems, but he focused only onconflict resolution. Furthermore.Cross dealt with static situationsrather than dynamic simulations.Similarly, the Advisor for the Intelligent Resolution of PredictedAircraft Conflicts (AIRPAC) system of C.A. Shively and K.Schwamb [7, 8) demonstrates arule system that handles juststatic-conflict situations. Morerecently, J.D. Reierson [9] reported on the use of a rule-basedsystem with dynamically simulated traffic to support the devel-
opment of the third and mostautonomous phase of AERA.
Solving static problems typically involves setting up a conflictbetween two or more aircraftwhose initial locations, velocities.and flight plans are given. Thecomputer program must thendetermine the best way to resolvethe conflict. The aircraft, however, are never moved orgiven theresulting ATC clearances. nor doother aircraft appear and causesubsequent conflicts at latertimes, nor are other ATC considerations typically taken into account. Although the staticmethod clearly captures someaspects ofATC problem solving. ithas the same artificial relation toATC as chess problems do to agame of chess.
Note. too. that there has beenextensive AI research in generalproblem-independent planningtechniques but it appears thatthis type of research is not applicable to theATC problem. (See thebox "Applying Artificial Intelligence Techniques to GeneralPlanning.")
The work reported in this article was influenced by earlierwork at Lincoln Laboratory on theElectronic Flight Rules (EFR) system (10). The system. which investigated the automation ofsome aspects ofATC, allowed pilots the freedom of visual flightrule (VFR) navigation but provided automated conflict resolution on an as-needed basis. Although the EFR conflict-resolution method was not an AI systemper se. the method's successfuluse of a cost function to evaluateproposed alternatives influencedour research.
538 The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
Spencer - Applying Artificial Intelligence Techniquesto Air n-aIfk Control Automation
References
1. RB. Wesson. Problem-SolvingwithSirnulation in the World ojanAir n-a1fic Controller. Ph.D. the-sis. Dept. ofComputer Sciences. 5.University of Texas (Austin. TX.1977).
2. P. Thorndyke. D. McArthur. andS. Cammarata. -Autopilot: ADistributed Planner for Air Fleet 6.Control. - Rand Corporation Report N-1731-ARPA (SantaMonica. CA. July 1981).
3. R Steeb. D.J. McArthur. S.J.Commarata. S. Narain. and W.D.Giarla. -Distributed Problem 7.Solving for Air Fleet Control:Framework and Implementation. - in Expert Systems: Techniques. Tools. and Applications.P. Klahr and D. Waterman. eds..(Addison-Wesley. Reading. MA. 8.1986). pp. 391-432.
4. RB. Wesson. K. Solomon. RSteeb. P. Thorndyke. and K.
Wescourt. -Scenarios for Evolution ofAir Traffic Control. - RandCorporation Report R-2698-FAA(Santa Monica. CA. November1981).RB. Wesson. -An AlternativeScenario for Air Traffic Control:Shared Control. - Rand Corporation Report P-6703 (SantaMonica. CA. Aug. 1981).S.E. Cross. Qualitative Reasoning in an Expert System Framework. Ph.D. thesis. CoordinatedScience Laboratory. University ofIllinois Report T-124 (UrbanaChampaign. IL. May 1983).C. Shively and K. Schwamb.-AlRPAC: Advisor for the Intelligent Resolution of Predicted Aircraft Conflicts. - The MITRE Corporation Report MT-84W164.(McLean. VA. 1984).C.A. Shively. -AlRPAC: Advisorfor Intelligent Resolution of Predicted Aircraft Conflicts. - inn-ansportation Research Circular
No. 310.Artifi.ciaLIntelligenceandthe Air n-affic Control System(Washington. DC. Dec. 1986).Report of a workshop held inOctober 1985. under the sponsorship of the Committee on Airfield and Airspace Capacity andDelay of the Transportation Research Board.
9. J.D. Reierson. presentation tothe workshop onArtificial Intelligence and Air Traffic Control.NASA Ames Research Center.Moffett Field. CA. 9-10 February1989. Published in AdvancedComputer TechnologyJor NAS. p.A-69. by the AdvancedATC Concepts Branch (ADS-120) of theFederalAviation Administration.
10. J.W. Andrews and W.M. Hollister. -Electronic Flight Rules: AnAlternative Separation Assurance Concept. - Project ReportATC-93. Lincoln Laboratory (31Dec_ 1980). FAA-RD-80-2.
human decision maker.Our project reqUired the development of a
working model of ATC planning and decisionmaking. This article presents the model alongwith justifications for choosing its particularstructure, and mentions some of the problemsthat result from using a fairly typical rule-basedprogramming system in an ATC environment.Then we discuss better alternatives that mightbe taken. Finally, a simple example extractedfrom one of the training scenarios with slightmodifications is given to illustrate the operationof the planner.
The work reported in this article addressesthe demonstration of automated problem solvinginanATC context. Two important subjectsthe role of such an automated decision makerwithin the overall ATC system, and how such asystem will communicate with controllers andpilots-were not studied.
An Approach to Automated ATe
What Controllers Do
A detailed account of the tasks and responsibilities of air traffic controllers is beyond the
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
scope of this article. For readers seeking suchinformation, Ref. 5 gives a very basic introduction to the operation oftheATC system; theATChandbook (6) contains the offiCial description ofthe controller's responsibilities; Ref. 7 providesa more engineering-oriented description; andthe Ainnan's Injonnation Manual (8) presentsATC procedures from a pilot's point of view. Aqualitative (if somewhat sensationalized) account of what a controller's job is like can befound in Ref. 9.
The following description focuses on thoseaspects of the controllers' job which must behandled by the automated system during thetwo training test problems. Specifically, the following functions will be covered:(1) coordinating with other sectors,(2) navigating aircraft,(3) issuing altitude clearances, and(4) maintaining aircraft separation.
Coordinating withothersectors. Each controller handles traffic in a volume of airspace calleda sector. Working together, controllers functionas a geographically distributed control system.The sectors may be adjacent either horizontallyacross the sector boundaries or vertically atspecified floor and ceiling altitudes. A given
539
Spencer - Applying ArtifICial Intelligence Techniquesto Air Traffic Control Automation
aircraft is handed off from sector to sector as itflies along its route. This handoff of controlresponsibility is the basic coordination functionbetween sectors. A controller must not allow anaircraft to leave his sector until the controller inthe next sector accepts responsibility for thevehicle.
Navigating aircraft. A pilot must file a flightplan as a prerequisite to entering the ATC system under Instrument Flight Rules (IFR). (UnderIFR, controllers are responsible for assuring safeseparation between aircraft. It is also possible tofly under Visual Flight Rules [VFRI, where thepilot assumes the responsibility for maintainingseparation from other aircraft. Under VFR, thepilot is not required to have extensive interaction with the ATC system.) As part ofan IFR flightplan, the pilot must precisely indicate whatroute of flight will be followed. This procedureconsists of chronologically listing either thespecific navigational fixes over which the aircraft will fly, or the segments ofairways that willbe used. (Airways are standard flight routes thatappear on aeronautical charts).
Before a flight commences, the ATC systemmust approve the route and issue a clearance tothe pilot. The pilot is expected to navigate theaircraft along the agreed-upon route. Controllers are not responsible for navigating aircraft,unless they elect to vector one off its route ontoa specified magnetic heading. In such a situation, controllers are responsible for the detailednavigation ofthe aircraft until it is put back ontoits original route or onto an alternative routethat the aircraft is capable of navigating on itsown. Vectoring is permitted only when the controller can observe the aircraft on radar and thuscan determine its precise location.
Issuing altitude clearances. Controllers aremuch more responsible for determining aircraftaltitudes. Although pilots indicate preferredcruise altitudes on their flight plans, they do notspecifY their vertical profiles as a function oftheir positions along the routes. Before they canchange from one assigned altitude to another,pilots must wait for altitude clearances fromcontrollers. In principle, this procedure is nodifferent from that ofreceiving route clearances.An aircraft, however, will typically receive only
540
one route clearance for the entire en routeportion of a flight, whereas several altitudeclearances are usually reqUired. In general, pilots prefer an immediate clearance to their filedcruise altitude at departure and another clearance to the airport approach altitude at a convenient distance from the destination airport.However, controllers often issue a series ofclearances so that aircraft must climb and descend instages.
Alternatively, a controller might issue whatamounts to an altitude profile in the form of aclearance with altitude restrictions that tell pilots to be at, at or above, or at or below certainaltitudes at certain points on their specifiedroutes. A series of such restrictions may beissued in one clearance that constrains thevertical profile over some portion of the route.
It is important to note that although thecontroller specifies the time at which an aircraftmay change its altitude and the altitudes thatare allowable, the controller does not control therate of climb or descent. Because the climb anddescent rates depend on the performance characteristics ofa particular aircraft, the pilot mustcontrol the rates. A controller can exercise somecontrol by means of altitude restrictions thatforce certain minimum rates, but a pilot is freeto reject such clearances if they are deemedunreasonable. From FAA tables [6], controllerscan gain some knowledge of the likely climb anddescent rates for most types of aircraft.
Maintaining aircraft separation. Safe distances between aircraft can be maintained either in a horizontal or vertical fashion. Horizontal-separation standards may be specified interms ofdistances or times; vertical separationsare always in terms ofaltitudes. The horizontalseparation standards that apply to a given situation are a function of a large number of variables [61. In the training scenarios used to testour computer program, the required horizontalseparation was 5 nautical miles (mni). Verticalseparation standards are a function of altitude.For aircraft flying below Flight Level 290 (FL 290,approximately 29,000 ftj, the minimum separation is 1,000 ft; for aircraft above FL 290, theminimum is 2,000 ft because of the lower accuracy of altimeters at high altitudes.
The LincoLn Laboratory JournaL. VoLume 2. Number 3 (1989)
To maintain the required minimum separations. controllers use a variety of methods.Under radar surveillance. aircraft may be vectored so that the vehicles remain horizontallyseparated. Under both radar and nonradarconditions. controllers may assign altitudes.speeds. and revised routes offlight to aircraft. oran aircraft might be delayed by holding it at aparticular point. To hold an aircraft. the controller instructs the pilot to perform either a standard 3600 tum or a racetrack-shaped pattern atsome point to which the pilot can reliably returnby using the aircraft's navigational equipment.
Spencer - Applying Artificial Intelligence Techniquesto Air TraffIC Control Automation
For a situation in which many aircraft must beheld at the same point. the controller separatesthe vehicles vertically. This procedure results ina holding stack of aircraft. Another delayingtactic. which can be performed only when anaircraft is under radar surveillance. is to vectorthe aircraft through path-stretching maneuverssuch as S-turns.
In general. vectors or altitude changes are thepreferred methods ofmaintaining aircraft separation. Because speed adjustments are greatlyrestricted by aircraft performance capabilities.such adjustments are used in en route ATe
Applying Artificial IntelligenceTechniques to General Planning
Reference 1 reviews the largebody ofresearch on applying artificial intelligence (AI) techniquesto general planning. Le.. planningthat is independent of specifictasks such as air traffic control(ATC). The report places the workin a common analytical framework. By and large. the approaches seem to have little applicability to real-life problems. Theapproaches focus on achievingsome well-defmed end point orgoal by determining the subgoalsnecessary to accomplish thatgoal. Steps are taken to ensurethat the methods used to achievethe subgoals do not interfere withone another. The state of the system being controlled is represented by logical assertions aboutthe relationships of various objects within the system. andchanges occur as discrete statechanges. For example. in thesimple toy-block stacking problems often used to illustrate thebehavior of these planners. ablock is represented as being on atable's surface at one step. andthen on top ofanother blockat thenext step. There is no representa-
tion for the continuous process ofmoving the block between the twostates.
One difficulty in applying thismethod toATC problems is that inATC there exist no particular endstates that need to be achieved.That is. ingeneral a large numberof possible future situations areacceptable. Another difficulty isthat the use of logical assertionsdoes not capture the continuousbehavior of physical systemssuch as aircraft in flight. andit also introduces a number ofartificial logical problems to thesystem.
Instead of using the above approaches as a foundation. theplanning methods deSCribed inour AI research derive more fromthe techniques used in gameplaying algorithms such as chessprograms. These games have thecommon characteristic that theycannot be thoroughly analyzed byworking backwards from a desired end point. This characteristic is due. in part. to the existenceofboth a large number ofacceptable end points and an even largernumber of possible intermediate
states leading to those end points.Thus. instead of working backwards. chess programs exploreforward from the current situation. The analysis is done by thesimulation of various possiblemoves and countermoves and theevaluation oftheir consequences.Numerous techniques thatstreamline the search processhave been developed. and muchwork has investigated the accurate evaluation of the hypothesized board positions. Althoughchess and most other games arerepresented by discrete boardstates. the same principles can beapplied to continuous systems byusing the appropriate simulations. This approach appears toavoid all of the problems of timeand change representation thatplag~e those systems which uselogical assertions orassumptionsofdiscrete-state behavior.
References
1. D. Chapman. -Planning for Conjunctive Goals.· ArtifICial Intelligence Laboratory Technical Report802. MIT (Nov. 1985).
The Lincoln Laboratory Journal, Volume 2, Number 3 (1989) 541
Spencer - Applying Artificial Intelligence Techniquesto Air Trajfic Control Automation
mainly for maintaining the desired separationsbetween aircraft in a sequence. Holding is apowerful technique, but because of the largedelay involved (aircraft need at least two minutes to execute one 3600 turn and longer for aracetrack-shaped pattern), it is not used exceptwhen other methods will not work.
A particularly difficult situation occurs nearairports where aircraft arriving along severalmerging routes must be formed into a singlestream of traffic. At such a location. organizingthe arriving aircraft into a final landing sequence is the major function of airport-approach controllers. The goal is to arrange theaircraft in a sequence with exactly the minimumallowable separation. thus achieving the maximum landing rate possible. Sequencing mayalso be needed for en route aircraft in areaswhere heavily traveled airways merge.
Why Is ATe Automation Difficult?
In automating the ATC decision-makingprocess, the central problem is that many factors might bear on a particular decision, andthere are usually a large number of actions orsequences of actions that could potentially betaken to resolve any problem. The factors varyfrom sector to sector (depending on the localgeography and traffic flows) and from situationto situation (depending on the exact configuration of aircraft in the sector).
Fig. 1-Alternatives for resolving aircraft conflicts.
542
For example. suppose two aircraft are flyingon courses that cross each other. The vehiclesare flying level at the same altitude, and theirspeeds and distances from the crossing pointare such that the separation standards discussed earlier are likely to be violated (Fig. 1). Ifonly this much information is available. possibleresolutions to the conflict might be to turn oneaircraft so that it travels behind the other. adjustthe altitude of one or both aircraft so they arevertically separated at the crossing point. ordelay one of the aircraft by either decreasing itsspeed or by implementing a delaying turn orholding maneuver. However. other considerations such as the following might invalidatesome of the above options:• There might be other aircraft in the vicinity
and the proposed maneuvers might createconflicts with them. In some cases, it isdesirable to resolve such secondary conflicts by maneuvering the secondary aircraft. It might be better in other instancesto resolve the original conflict anotherway.
• Severe weather conditions might prohibitsome of the proposed options. Also, an aircraft's proximity to the ground, mountains, or other physical obstructions mightforbid certain maneuvers.
• If a maneuver forces an aircraft to cross orcome close to a sector boundary, the controller is burdened with the additionalwork ofhaving to coordinate the maneuverwith the controller of the adjacent sector.This solution. however, might be acceptable ifother options are even less desirable.
• Aircraft that are close to their maximumflying altitude will be unable to climb farther. Furthermore, aircraft without pressurized cabins or oxygen masks are forbidden to climb above 10,000 ft.
• Certain maneuvers, either by themselvesor by the secondary conflicts they cause,might result in undue delays.
• Requiring ajet to fly long distances at lowaltitudes wastes fuel.
• Itis inefficient to force a climbing aircraft todescend or a descending aircraft to climb.
• If the descent of an arriving aircraft is delayed, the vehicle might not have enough
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
Spencer - Applying Artificial Intelligence Techniquesto Air TrafflC Control Automation
Combined Evaluation.-------.....~I
PlanGenerator
Current
Plan
Combine
Plan Critics
IndividualEvaluations
SuggestedChanges Proposed Plan
AircraftState
Prediction
Plan + Predictions
Fig. 2-Plan evaluation and modification flowchart.
time to reach the appropriate altitude reqUired to approach the airport for landingpurposes.
Hundreds of such considerations can affect theacceptability ofa particular solution. Whether aparticular consideration is applicable dependson the context in which the situation occurs,e.g., the geographic location and the existence ofother aircraft in the vicinity.
It is difficult to devise a software structurethat can accommodate all applicable considerations, evaluate all possible solutions, make thenecessary trade-offs among the solutions, and,at the same time, be modular, maintainable,adaptable to local-site considerations, and fastenough to keep up with the dynamic nature ofATC.
Problems often arise when an automaticdecision-making system is integrated withhumans. Specifically, what recourse should betaken if human controllers disagree with thedecisions of the expert system? One approachwould be to make the automated system sointelligent that its plans and decisions werealways understood by and usually acceptable tohuman controllers. A model for this interaction
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
would be the current practice of having controllers work in teams of two or three to control asector. In the team, one person is the primarycontroller and the others are responsible forrelieving him of certain auxiliary functions.Communication among the team membersdepends on each one's comprehension of theATC process at a very high level so that everyoneunderstands everybody else's actions withoutany need for long explanations. However, forthe automated system to achieve such a highlevel of competence, it would have to take intoaccount all of the above-mentioned considerations. As stated earlier, the current clericalsystems avoid this problem, as they merelysupply additional data to controllers. That is,the current systems do not have to integrate thedata into a final decision.
A Software Structure forATC Problem Solving
Figure 2 diagrams a software structure thatappears to address the above concerns. Thestructure also seems capable of solving problems in a way that is intuitive and understand-
543
Spencer - ApplyingArtiftcial Intelligence Techniquesto Air Traffic Control Automation
able to human controllers. In the figure, a plangenerator produces plans that detail how theautomated controller will handle the currenttraffic situation for the specified sector. Each ofthe sector plans contains a set of individualaircraft plans that specify the routes of flight forthe aircraft and the ATe clearances that areplanned at particular points along these routes.Clearances might be altitude changes, speedchanges, vectors, or holding commands. Fromthe above information, it is possible to projectwithin an approximate range the future horizontal positions and altitudes of the aircraft.
Given a plan and projections of the futurepositions and altitudes of aircraft, it is possibleto write plan critics. A plan critic is an independent software module that is responsible forlooking for a particular type of undesirablefeature or consequence of a plan. Each plancritic produces a score that represents themodule's evaluation of the plan from the module's particular point ofview. In our system, thehigher the score, the more severe are the problems with the given plan.
The individual scores are then weighted andcombined by a simple summation into an overallscore for the plan. The resulting score is fed backto the plan generator, which uses the score torank the given plan against other possible plans.It is this combining function that allows thesystem to make trade-offs among the variousconsiderations represented by the individualcritics (10). The individual scores are weightedto give the correct trade-offs among the variousproblems so that the system ranks plans in adesired order. In the current implementation ofour system, the weights are adjusted byempirical analysis. The weight-adjusting process(which reflects learning) could perhaps be automated with techniques that neural networkresearchers are currently exploring. It mightalso be possible to determine the weights in ananalytic way, e.g., by assigning to each type ofproblem an estimated cost in dollars.
The overall effect of the scoring process is thecreation of a function that takes a plan as itsargument and returns a number. Given thathigher scores denote less acceptable plans, theplan generator's goal is to find a plan with as
544
small a score as possible. Unfortunately, analytic techniques cannot be applied because thefunction is highly nonlinear. Also, the exactfunction will change as the system evolves andnew plan critics are added. Thus an efficientsearch process is needed. Undirected searchcould become very expensive, as could searchfor the true optimum plan.
A! systems typically require the derivation ofan efficient search algorithm. A common approach is to apply knowledge of the task domain(in our case, air traffic control) to restrict thesearch process so that good, albeit suboptimal,solutions are found quickly. In the approachproposed in this article, the same plan criticsthat score the plans can often provide suggestions as to how a plan could be improved from alocal point of view. However, a particular plancritic is not able to determine the effect of itssuggestions on the overall score because thecritic purposely has no knowledge of other critics. Thus critics merely feed their suggestions tothe plan generator, which generates new planscorresponding to the suggestions and passesthe plans back to the critics for evaluation.
In searching for a solution, the plan generatorexplores some portion of a search tree in whichthe tree's root is the original plan. In the tree,child nodes represent plans that have beenderived from their parent node's plan. The plangenerator implements a particular search strategy to find the branch of the search tree thatshould be pursued next. The search strategyselects some already evaluated plan and itscorresponding suggestions, and generatesmodified versions of the plan according to thesuggestions. The search strategy may then electto pursue one of the new branches further or tofollow some other older branch that appearsmore promising. A stopping criterion that determines when a plan is acceptable must be defined. Acceptable plans are passed to that partof the system which implements the currentplan. The current plan remains in effect until theoccurrence ofany event that increases the plan'sscore. An example would be the appearance of anew aircraftwhose individual plan conflicts withthe individual plans of some of the aircraft thatare already known to the system. This type of
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
situation initiates a new round of searching.
Limitations of the CurrentSystem Implementation
Our expert system is implemented in a combination ofLISP and YAPS (Yet Another Production System) [II, 12J. As its name implies, YAPSis a fairly standard rule-based system that usesforward chaining [13J. YAPS allows arbitraryLISP functions to be used both on the right-handside of rules (the then part ofan if--then rule) andas tests on the left-hand side (the ifpart). Thisflexibility contrasts with many other rule-basedsystems whose computational capabilities arerestricted to the functionality provided by theirlanguage designers. The implementation ofYAPS used in our expert system was originallyconverted from Franz LISP to Zetalisp at MIT.Subsequent modifications at Lincoln Laboratory improved the implementation's executionspeed and debugging facilities.
The general architecture described aboveappears to have the power and flexibility neededfor automation of the ATC decision-makingprocess. The current implementation, however,is limited in several respects and much furtherwork is required to develop it fully. But even withthe limitations, the expert system was able tohandle the first two training problems of theBoston Air Route Traffic Control Center(ARTCC).
The current system has a very simple searchprocedure that searches only one level deep inthe tree. The one-level restriction results in thefollowing limitation. When the current plan hasa problem, the plan generator will look only atthose alternatives which are immediately derived from that plan. If some of the alternativesalso have problems (such as secondary conflictscaused by maneuvers that resolved the initialproblem), then the alternatives will receive highscores and will most likely be rejected. Note thatif the system were to search deeper to generateresolutions to secondary problems, a plan thatwas better than any of the first-level plans mightbe discovered.
Another limitation of the current system is itsmeager number of plan critics. The imple-
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
Spencer - ApplyingArtiflCiallntelligence Techniquesto Air 1Taffic Control Automation
mented critics identifY conflicts between aircraft, detect when an incoming aircraft will notreach the appropriate terminal altitude in time,and complain if an overflight is cleared too farfrom its requested cruise altitude. For the current system to evolve into a competent air trafficcontroller, we must develop all of the neededcritics and adjust their scores so that they reflectrankings that human controllers would consider appropriate. In the architecture of oursystem, the plan critics and the scoring functionare the primary repositories of ATC knowledge.
The current system is also constrained by itsplanning and execution functions, which have avery limited repertoire of actions that are available to resolve problems. Thus most conflicts areresolved by altitude changes. The only exceptions are situations in which the routes of twoaircraft merge and a conflict occurs either at themerge point or at the time when one of theaircraft overtakes the other. These situationsare resolved by directing one of the aircraft tomake a 3600 right tum. Another weakness ofthecurrent system is that it does not check to makesure that the airspace in which this tum will bemade is free of conflicts.
Because of the limited number ofoptions andthe limited depth of search, the plan generatorcan try all of the suggested solutions to a givenconflict in a reasonable amount of time. Ultimately the plan generatorwill have to use a moreselective strategy that searches only the mostpromising branches of a search tree. In such astrategy, the best plan (as defined by the system's evaluation function) might be missedunder certain circumstances. The nature ofATC, however, is such that there are typicallymany acceptable solutions to anyone problem,and finding the best solution is usually unnecessary. It is also possible in the future that thoseimplementations which are optimized for speedand/or implementations that use special hardware might be able to conduct a full search in
real time.
A Development Environmentfor the ATC Expert System
To test the automated controller with differ-
545
Spencer - ApplyingArtiftcial Intelligence Techniquesto AirTraIfk Control Automation
TrafficEnvironment Knowledge (Scenario
FileFile File
+ !,Position Reports
,Traffic
Flight Strip Data Automated
Simulation ATC-Aircraft Messages Traffic
Sector-Sector Messaqes Controller
Simulation lDisplay ControlControl Observer's
Interface
+Traffic r Observer's 'IDisplay Display
j
Fig. 3-Development environment of the ATC expert system.
ent traffic scenarios, we implemented a development and testing environment (Fig. 3) on aSymbolics 3670 computer. A set of interfaces
provides the expert system with the same information that a real controller receives. Suchinformation includes radar position reports;
546
AC49 OC9/A 290 G461 H61 (BOA 180 10) 0135 JFK' ./. (BOA 180 10) CAM PLB ./. YUl'1.4707 C141/A 350 G480 T480 NOPAl 0136 BGSF' ./. NOPAl ALB J37 JFK ./. WAI' e1 :42:38TW41 B707/A 370 G450 T450 lOXXE 0137 OAO' ./. lOXXE J82 ALB GOM V431 BOS'AC97 OC9/A 330 G490 T490 (BOA 180 10) 0138 JFK' ./. (BOA 180 10) CAM PLB ./. YUl'
InitializeReal-timeFaat-tlmeSlow-time
BDmlDlaabled L1atenlng~
Controller Debug InfoTraffic Situation Display
olor Screen91:36:~4 Other Seetor Athens Request Handorf AC49 "!!mh3"1"',! .aYAg"91 :36:44 Athens Other Sector Accept H.ndor, AC4991:36:45 AC49 Athens Radio Checkin91 :36:45 Athens AC49 Cl iP'lb And Meintain 299 White-an-black91:37:24 Other Sector Athena: Request Handaff "797 h91:37:24 Athens Other Sector Accept Handorf "797 0 5 II! 15 20 25 INF91 :37:25 "197 Athens Radio Check1"91:38:14 Other Sector Athens Request Handaff Tu4191:38:14 Rthens Oth.,.. Sector Accept Handorf Tu41 Add Airway Map91:38:15 TU41 Athens Radio Checkin Remove Sec'tor Map91:39:44 Other 6ector Athens Request Handorr AC97 Display A91:39:44 Athens Other 6ector Accept Handorr AC97 Display B91 :39:45 AC97 Athens Radio Checkin Display C91:39:45 Athens AC97 Cli"b And "ointain 299
Display 0New Display
Env 1ronnent ril e na"e ~derau 1t .tc.in: scenari os; athens-envi ron"ent) AT CSI t1: SCENARIOS; AT HENS-ENVIRONtlENTScenario rOe na"e (default atcsi"lscenar108Jathens-l) cro.sing-conflict-.cenar1oController knouledge-b••• fil. (def.ul t .tc.i"~sY8te"; athens-knouledg.) ATC6It1~6't6TEI1;ATHENS-KNOlollEDGERecording rile - END ror none [Hon.]
Fig. 4-0bserver's display.
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
Spencer - Applying Artificial Intelligence Techniquesto Air Traffic Control Automation
flight-strip data, which give the routes of flightsand their desired cruise altitudes and speeds;radio messages to and from aircraft; and. forcoordination purposes. messages to and fromadjacent sector controllers. Except for positionreports. information flowing across the interfaces is displayed on a monochrome screen.which allows observation of the system's performance. The display (Fig. 4), which providesmenus that allow an operator to control thesystem's operation, contains three windows:one for displaying flight-strip information. another for displaying messages between the system and pilots or adjacent sector controllers.and a third for entering input parameters fromthe keyboard.
A color screen displays aircraft positionsalong with airways. navigational aids, airports.and sector boundaries. Figure 5 shows a monochrome version of this display. In the figure.aircraft positions are represented by dots surrounded by 5-nmi-diameter circles. which provide distance references. The screen containstrack histories (information detailing the previ-
ous positions ofaircraft). whose time lengths arecontrolled from the observer's display. Eachaircraft symbol is accompanied by a data tagsimilar to those on standard ATC displays. Adata tag gives an aircraft'sflightidentity, clearedaltitude, actual altitude (if different from thecleared altitude), and ground speed.
The flight-strip information is mouse-sensitive; Le., flight strips act as menu items. Whenthe flight strip of an aircraft is selected. a firstlevel menu appears and allows the operator to doseveral things: issue ATC commands to theaircraft. change the position of the data tagrelative to the aircraft symbol, perform certainactions as if piloting the aircraft. or performactions relating to intersector handoffofaircraftresponsibility. When the ATC-command or pilot-action mode is selected, a second-level menuappears and shows the clearances or actionsaccepted by the simulated aircraft.
A major component of this development environment is a traffic simulator based on a systemdeveloped by Antonio Elias and John Pararas atthe MIT Flight Transportation Laboratory. The
81:42:38
o
o
In t. zeR••I-tlm.F••t-tlm.Slow-tim.
IlD!m
Whlt.-on-bl.ck
Color Scre.n'M,milIY'!;·!"'''·;A;'!
Flight StrlpaIM••••g••. Controller Oeb'1lnro
11FI1,,*iiUriimTj_ IIi!dItW
o 5 15 20 25 INF
REVER
Ao BOS
'J PVD
PUT
o
oOR~
I••bl.d L1.t.nlng
oHFD
oCAM
CANAN
A
LAMED RVMMYe-- M1B1ENE
A A0359C
4B9
o ALB
AC49
IGN O~::C
oHUO
HNK
UCA T~41
319C459
........lo
RKA
Dis 1e E
Fig. 5-Traffic-situation display.
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989) 547
Spencer - Applying ArtifICial Intelligence Techniquesto Air Traffic Control Automation
resulting simulator is driven by a traffic scenariofile that specifies the same infonnation for eachaircraft as would be found on a flight plan,specifically, infonnation regarding the type ofaircraft and the vehicle's flight route, cruisealtitude, and cruise speed. Other initializationinfonnation such as the altitude of the aircraftcan be provided, and there is also a mechanismfor specifying the time when certain ATC actionsare scheduled to occur. Such pre-scripted actions include issuing altitude clearances thataircraft would receive in adjacent sectors, andinitiating handoffs. This capability mimics asimilar time-based function in the simulatorscurrently in use at en route centers. In fact, theentire scenario file is a fairly direct translation ofthe scenario card decks used at the centers.Data were obtained for the first 10 of the 18problems that the Boston en route center wasusing in late 1983. The two simplest problemswere turned into scenario files for use by theexpert system.
In addition to scenario files, the simulatorand the expert system use an environment file,which provides map infonnation about the region of interest. The map infonnation includesthe positions of navigational aids and airports,the routes of all airways used in the scenarios,and the boundaries of ATC sectors.
Our development environment serves twopurposes. As indicated above, it allows the testing of the expert system on lifelike problemsdUring which developers and human expert airtraffic controllers can observe the system'sbehavior. It is also possible to disconnect thedevelopment environment from the expert system and allow the observer to control the trafficthrough the menus mentioned above. In thisway it is possible to see how a controller wouldhandle a certain situation, or to allow developers to try out certain control approaches manually before programming them into the expertsystem.
An Example
In this section we present a simple examplethat demonstrates some of the system's concepts and limitations discussed earlier. The
548
example is abstracted from the first of the 18Boston ARfCC training problems. All of theproblems were set in a high-altitude (i.e., above18,000 ft) sector called the Athens sector. Figure6 shows the initial traffic situation. The onlychange made to the original scenario was toadjust the timing of the aircraft labeled M707 sothat it would conflict with 1W41 over Albany(ALB), N.Y. Flight M707 is an overflight; 1W41is an arrival for Boston (BOS); and AC49, a departure from John F. Kennedy InternationalAirport in New York City, is headed for Montreal.
The upper portion ofFig. 7 shows the altitudeof 1W41 as a function of the distance traveledalong the aircraft's path. 1W41 enters from theleft and flies level at its cruise altitude. Justbefore Albany, however, the planner intends toissue a descent clearance. From that point on,1W41's altitude is shown as a band that represents the planner's uncertainty about the aircraft's descent rate. The upper line of the bandindicates the minimum descent rate, the lowerline the maximum rate, and the middle line thenominal rate. The locations of the rectangUlarboxes labeled M707 and AC49 show where thetwo aircraft will cross 1W41's path. The boxesindicate a 2,OOO-ft altitude separation and5-nmi horizontal separation on both sides ofthe planned positions ofthese aircraft. Thus potential conflicts between the aircraft are denoted by areas where 1W41's path intersectsthe boxes. If1W41 is predicted to be in a box atthe same time that the box's aircraft is expectedto be there, a conflict is declared.
The lower portion of Fig. 7 shows the corresponding state ofthe planner's search tree. Theplanner's initial state is shown at a time prior to1W41's appearance. Thus no problem has beendetected yet and the initial state has a low scoreof zero. The planner goes from that state to astate in which 1W41 appears and the conflictwith M707 is discovered. The second state,therefore, has a high (i.e., an undesirable) score.The planner then tries to reduce the score. Notethat only 1W41's conflict with M707 has beendetected at this point.
We designed our rule-based system so thatthe discovery of a conflict leads immediately toan investigation that tries to resolve the conflict.
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
TW41
o HNK
OPVD
Fig. 6-lnitial-conflict situation.
Thus, in the example, the other conflict withAC49 will not be discovered until after the M707conflict is resolved. A better implementationwould be to fmd all of a given plan's conflictsfIrst, and then prioritize the conflicts so that theycan be solved in order of their importance.Alternatively, a sophisticated system might fIndcases in which several conflicts could be re-
Spencer - Applying Artificial Intelligence Techniquesto Air Traffu; Control Automation
solved by a single action. These types of advanced algorithms are difficult to implement ina rule-based system like YAPS.
In the example, the critic that detects aircraftcrossing conflicts proposes four possible solutions: 1W41 can be restricted to pass eitherbelow or above M707, or M707 can be restrictedto pass either below or above 1W41. For the fIrstcase in which 1W41 is constrained below M707,Fig. 8 shows the planner's search state and thecorresponding altitude plan. The heavy blackline below M707 represents the restriction that1W41 must be at or below 33,000 ft over theportion of path that is denoted by the heavyline's length. Upon evaluating this new plan, thesystem finds a subsequent conflict between1W41 and AC49. Consequently, this plan receives a high score.
The example demonstrates the search limitation of our current system. Because the systemsearches only one level deep in the search tree,the secondary conflict between 1W41 and AC49could not be solved, for example, by starting1W41's descent earlier so that the aircraftwould
INITIAL CONFLICT VERTICAL PROJECTIONS
400
TW41
:2g 350
~(l)'0
E 300
«
250
50 60 70 80 90 100Distance along Path (nmi)
PLANNING STATE
~41
o 10Appears
C49
110 120
Fig. 7-lnitial-conflict vertical projection (top figure) and planning state (bottom figure).
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989) 549
Spencer - Applying Artificial Intelligence Techniquesto Air Traffic Control Automation
RESTRICTING TW41 BELOW M707
400TW41
120o
AC49
M707
70 80 90 100Distance along Path (nmi)
60
250 '--__..I....-__...I-__...J....__---I...__----I._----:
50
E'g 350
Q)"0.~ 300<i:
PLANNING STATE
o ~TW.;..;...;.4.;..1_...Appears
TW41 Below
Fig. 8-Examining the consequences of restricting TW41 below M707.
fly below both M707 and AC49. Instead, thesystem altogether rejected the alternative ofrestricting 1W41 below M707.
Figure 9 shows the remainder of the searchtree for the first conflict and the selected altitudeprofile. The selected plan holds 1W41 high untilthe aircraft has passed over M707. The systemhas determined that1W41 still has enough timeto descend into Boston, albeit with less marginthan in the original plan. The conflict-detectioncritic ignores the comer where 1W41's maximum descent rate might cause it to enter AC49'sprotected airspace. As currently implemented,the critic is more concerned with the nominalprojection than with the maximum and minimum projections. However, a more advancedplanner might treat this situation as a secondary conflict that needs to be resolved by restricting 1W41's altitude to be above that ofAC49.
In the example, the other two options (M707restricted above 1W41 and M707 restrictedbelow 1W41) for resolving the original problemreceive scores equal to the option of restricting1W41 above M707. The planner randomly se-
lects one solution. However, forcing this choicewould be better because, in general, it is preferable not to disturb an overflight. This additionalcriterion could be implemented by the creationof a critic that penalizes those plans whichinterfere with overflights.
Now that the planner has solved the firstconflict, another aircraft appears (Fig. 10). AC97has the same route as AC49 but the aircraft hasfiled for a slightly higher cruise altitude. Figure11 shows the resulting altitude plan and planning state. The newly proposed plan for 1W41now conflicts with the initial plan for AC97. Theplanner tries to resolve this conflict by using thesame four options as before: 1W41 above,1W41below, AC97 above, and AC97 below. The basisfor these options is the current plan, which is theresult of resolving the first conflict describedabove, and the addition of the plan for AC97.
If 1W41 is held above AC97, 1W41 will nothave enough time to descend to 8,000 ftfor entryinto the Boston terminal area. Thus that particular alternative receives a high score. Whenthe planner evaluates the effect of trying torestrict 1W41 below AC97, the state-prediction
550 The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
Spencer - Applying Artificial Intelligence Techniquesto AirTrajfk Control Automation
RESTRICTING TW41 ABOVE M707
400
TW41
£g 350
~OJ"0
'S 300
«
250
50 60
M707
70 80 90 100Distance along Path (nmi)
PLANNING STATE
110 120
o TW41
Appears
TW41 Above
M707 Below
Fig. 9-Resolution of the initial conflict.
routines determine that because 1W41's descent rate is limited, the aircraft cannot go bothabove M707 and below AC97. Therefore, theconflict is not resolved and this option alsoreceives a high score.
Figure 12 shows the final planning state andthe selected altitude plan. Note that the scoredue to 1W41's inability to reach the properaltitude for terminal-area entry is not as bad asthe score for an option that contains a confliCt.The planner chooses the plan that restrictsAC97 below 1W41 until AC97 is past the pointof conflict, and then clears AC97 to its desiredcruise altitude. This alternative has the samescore as the alternative of restricting AC97above 1W41 and, once again, the final solutionis arbitrarily chosen. As in the previous case,there are reasons that the chosen alternative issuperior; additional critics could be written torepresent these criteria in order to force thesuperior solution.
Figures 4 and 5 show the development system's displays approximately 2 min after thesecond conflict is resolved. Note that AC97 hasbeen cleared to 29,000 ft rather than its re-
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
quested cruise altitude of 33,000 ft.
Conclusion
The system described in this article wascapable of successfully perfOrming basic ATCand conflict-resolution tasks for the easiest two
TW41
o HNK
OPVD
AC49o BDR
1AC97
Fig. 1o-Second-conflict situation.
551
Spencer - Applying Artiftcial Intelligence Techniquesto Air Traffic Control Automation
of the Boston ARTCC training scenarios. Thesystem, which operated in real time for the firstscenario, used a Symbolics 3670 computer thatalso ran the simulation and display programs.The system's limited knowledge base restrictedthe available number of problem-resolutionactions and kept the system's behavior frombeing robust, or truly expert. However, for thesetwo scenarios and some smaller scenarios derived from them, the system maintained properaircraft separation, cleared aircraft to theirproper altitudes, and coordinated handoffs withadjacent sectors.
The major conclusions of this study are thefollowing:• The system's overall planning architecture
appears capable of performing automatedATC. The architecture satisfies the basicfunctional requirement of being able to incorporate ATC knowledge in a modularfashion. This modularity allows the systemto make trade-offs among the possiblyconflicting recommendations ofthe knowledge sources. The architecture also en-
abIes the search for good (but not necessarily optimal) solutions in a reasonableamount of time.
• Further work is needed to expand the system's search process and evaluationmechanism. In particular, multilevelsearch and a well-defined method for developing the evaluation weights appearnecessary. The similarity ofthe basic planning mechanisms of our system to thoseused in automated chess programs impliesthat useful gUidance might be obtainedfrom recent advances in that area. Thestructural similarity of the evaluationmechanism to neural networks suggestsan adaptive learning approach to theweight-setting process.
• An extensive amount of work is needed,both to capture the necessary ATC knowledge in the plan critics, and to test thesystem on a large enough sample of problems to ensure error-free performance.One way of accomplishing the above whileobtaining some benefits dUring the devel-
SECOND CONFLICT VERTICAL PROJECTION
400TW41
£"g 350
~Q)"0~ 300'';:::
~
25050 60
M707
m 00 00 100Distance along Path (nmi)
PLANNING STATE
AC49
110 120
552
Fig. 11-Second-conflict vertical projection (top figure) and planning state (bottom figure).
The Lincoln Laboratory Journal. Volume 2. Number 3 (l989)
Spencer - Applying Artificial Intelligence Techniquesto Air Traffic Control Automation
RESOLUTION OF SECOND CONFLICT
400
£g 300
AC97
TW41
100 L...-__..L..-__-'----__--'-__---'-__----'-__----J'--_----'
o 20 40 60 80 100Distance along Path (nmi)
PLANNING STATE
AC97
Appears
120 140
TW41Below
AC97Above
Fig. 12-Resolution of the second conflict.
opment peIiod would be to use the systemas a training aid. By doing so, we could usethe large number of already developedtraining scenaIios.
• Simple forward-chaining rule-basedmechanisms are not suitable for systemswith both nonstatic data and the need torepresent mutually exclusive alternativeswhile searching for solutions. In suchsystems, two types ofmechanisms, one fordeleting deIived facts that have becomeinvalid and the other for keeping alternative plans separate, must be explicitlyprogrammed. This requirement makes therules more complex. Assumption-basedtruth-maintenance systems [14] appear toprovide these types ofmechanisms as partof their basic design while, at the sametime, allowing for the modular representation ofknowledge through the use ofrules.Such systems need to be tested in an ATecontext to determine if their expectedbenefits can be achieved while real-time
The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)
performance is maintained.• When a problem involves large numbers of
similar facts, the standard ways of implementingrule-based systems are subject tocombinatoIial explosions of rule-and-factmatches in the system's working memory.The repercussions of such explosions canbe ameliorated with coding tIicks, but thissolution places a great burden on programmers and makes the coding less comprehensible. The use of coding tIicks alsoleads to system behavior in which a smallchange in a rule or the addition of a newrule greatly affects performance. This effect bIings into question the suitability ofarule-based approach for situations inwhich real-time performance is required.In rule-based approaches, there appears tobe a trade-off between the potential modulaIity and compactness of a system, andthe system's speed of execution. However,this trade-off may not be fundamental,and improved automated optimization
553
Spencer - Applying ArtifIcial Intelligence Techniquesto Air Traffic Control Automation
techniques may result in a system that ismodular. compact. and capable of realtime operation.
Acknowledgments
The author thanks Don Underwood. who didmuch ofthe programming to develop the systemdescribed in this article; and Steve Olson. whodid most of the YAPS start-up and modificationwork. The author thanks Steve Campbell. whoalso was the source of many useful ideas andcomments.
This work was sponsored by the FederalAviation Administration.
References1. Special issue on the FAA's Advanced Automation Sys
tem (AAS). Computer 20 (Feb. 1987).2. AH. Gisch and B.C. Zimmennan. "AERA-Toward
Greater En Route Air Traffic Control Automation." TheMITRE Corporation Report MP-86W29 (Mclean. VA.Oct. 1986).
3. J.A. Kingsbury. "AnAERAfor This Century." TheMITRECorporation Report MP-86W28 (Mclean. VA, Oct.1986).
DAVID A SPENCER is asenior staff member of theSystem Design and Evaluation Group at Lincoln Laboratory. His focus of research
has been in computer software systems related to air trafficcontrol. Dave received a B.S.E. in electrical engineering fromPrinceton University. and an E.E. and S.M. in electrical engineering and computer science from MIT. where he completed his master's thesis at the university's Artificial Intelligence Laboratory.
554
4. D.A Spencer. J.W. Andrews. and J.D. Welch. "AnExperimental Examination of the Benefits ofImprovedTerminal Air Traffic Control Planning." LincolnLaboratory Joumal 2, 527 (1989).
5. G.A. Gilbert. Air TraffIC Control: The Uncrowded Sky(Smithsonian Institution Press. Washington. DC.1973).
6. Air Traffic Control, Order 71 10.65. U.S. Department ofTransportation. Federal Aviation Administration(Washington. DC. updated periodically).
7. H. Ammennan. C. Fligg Jr.. W. Pieser. G. Jones. K.Tischer. and G. Kloster. "En Route/Tenninal ATCOperations Concept." ComputerTechnology Associates.Inc.. Report (Englewood. CO. 28 Oct. 1983) DOTIFAAIAP-83-16.
8. Airman's Injormation Manual, U.S. Department ofTransportation. Federal Aviation Administration(Washington. DC. updated periodically).
9. D. Biggs. Pressure Cooker (W.W. Norton. New York.1979).
10. J.W. Andrews and W.M. Hollister. "Electronic FlightRules: An AltemativeSeparation Assurance Concept."Project Report ATC-93. Lincoln Labor~tory (31 Dec.1980). FAA-RD-80-2.
11. E.M. Allen. "YAPS: Yet Another Production System."Department of Computer Science. University of Maryland. ReportTR-1146(Feb. 1982).
12. E.M. Allen. "YAPS: A Production Rule System MeetsObjects." Proc. oj the Ntl. Conj. on Artificial IntelligenceAAAI-83. Washington. DC. 22-26 Aug. 1983. p. 5.
13. P.H. Winston. ArtificialIntelligence. 2nd ed. (AddisonWesley. Reading. MA. 1984).
14. J. de Kleer. "An Assumption-Based TMS." ArtificialIntelligence 28, 127 (1986).
TIle Lincoln Laboratory Journal, Volume 2. Number 3 (1989)