Download pdf of this article

Distributed Cooperative Problem-Solving in the Air Traffic Management System

Abstract | The air traffic management system in the U.S. is an example of a distributed problem-solving system. It has elements of both cooperative and competitive problem-solving. This system includes complex organizations such as airline operations centers, the air traffic control systems command center and traffic management units at enroute centers and TRACONs, all of which have a major focus on strategic decision-making. It also includes organizations and individuals concerned with tactical decision making (such as air traffic controllers and pilots). The architecture for this system has evolved over time to rely heavily on the distribution of tasks and control authority in order to keep cognitive complexity manageable for any one individual. Currently, major changes are being made in this architecture, especially with respect to the locus of control, in an effort to improve efficiency and safety. This paper discusses some of these changes from the perspective of system complexity and illustrates the need to maintain a clear understanding of what is required to assure a high level of performance when alternative system architectures and task decompositions are developed.

Introduction | In the design of complex systems, there is often a tension between a desire to achieve solutions that are globally "optimal" and the limitations imposed by the cognitive complexity of determining such "optimal" solutions. These limitations can arise either:

•   Because the individual operators who are trying to make the system function could not deal with the cognitive complexity of the task if they were to fully consider all of the relevant goals and data in arriving at an "optimal" solution; or
•   Because the designers of technological tools intended to make it possible to overcome the cognitive limitations of human operators are themselves unable to fully model the true complexity of the system.

Because of these two types of limitations, most real system designs therefore rely on simplifications that allow the system to perform well, without trying to determine and implement "optimal" solutions. One common approach is to decompose the task of managing the overall system into subtasks, and then assigning these subtasks to separate individuals. The hope is that there is sufficient independence among these subtasks, so that when each subtask alone is performed well, the combined effects produce good performance for the system as a whole. Furthermore, because few systems are actually decomposable into fully independent subtasks, it is also hoped that these individuals will interact with one another as needed when the solutions to their various subtasks in fact interact.

Below, these considerations about the design of complex systems are discussed in the context of the Air Traffic Management (ATM) system. Particular emphasis is placed on the need to understand how task decomposition influences performance, and how attempts to move toward designs intended to integrate decision making and thus to produce closer to "optimal" performance must consider the need to adequately support new task decompositions if they are to be fully effective.

An Example of Task Decomposition in Air Traffic Management | As a specific example of task decomposition, consider the following scenario. In order to reduce cognitive complexity, the overall task of selecting safe routes of flight and of operating these flights is currently decomposed such that each of the participants (pilots, controllers, dispatchers and traffic managers) has only partial information. In particular, within the current air traffic management system, tactical decisions are made by flight crews and controllers without always having the information necessary to develop the same big picture about weather system developments available to dispatchers and traffic managers. Because this decomposition is not always adequate, controllers sometimes request reroutes for flights which do not have sufficient fuel for the proposed reroute. Thus, in cases involving "significant" reroutes, the flight crew should bring the dispatcher back into the loop to ensure that the big picture has been adequately considered. Although this distribution of information and responsibilities generally affords an efficient operation, it is susceptible to occasional errors due to false assumptions about "what the other guy has already considered" or due to incorrect assessments of whether a particular change in route is "significant".

Details of the Scenario | As an example, consider an actual incident involving a Boeing 727-200 flying from Dallas/Ft. Worth to Miami. As part of his job, the dispatcher responsible for this aircraft was required to provide the pilot in command information regarding any hazardous enroute weather. In this case, the dispatcher noted a line of thunderstorms that he felt potentially jeopardized the safety of the flight and issued a reroute to the aircraft, with the Captain's concurrence. During this process, the Captain was briefed on the situation. That reroute was coordinated with ATC and approved, but as the flight progressed along its refiled route of flight, the receiving Center rejected the reroute and put the airplane back on its originally filed route of flight. As a result, the aircraft became trapped south of the line of weather.

More specifically, as the aircraft was going across the Florida panhandle, there was a line of thunderstorms from the Tampa Bay area southeastward down to the Miami/Ft. Lauderdale area. At that point, the dispatcher contacted the captain, briefed him on the enroute weather conditions, and recommended a reroute taking the aircraft direct Ormond Beach and then down the east coast of Florida into the Miami airport from the northeast, ahead of the weather. The Captain concurred with the reroute and contacted the appropriate Jacksonville Center frequency to coordinate the reroute. The reroute was approved. The aircraft made a turn to the east and was proceeding direct Ormond Beach on the Florida East Coast. At the point where there was a hand-off made from one controlling Center sector to the next Center, the receiving Center sector advised the Captain that, due to traffic along the east coast of Florida, they would not be able to accommodate the reroute and that the aircraft would have to return to the originally filed route of flight. The aircraft made a fairly abrupt turn back to the southwest, got offshore along the west coast of Florida and proceeded down toward the Ft. Myers area. Furthermore, the aircraft was slowed to 180 knots due to traffic, increasing fuel burn.

At that time, the line of thunderstorms was sinking to the southeast, moving down toward Miami/Ft. Lauderdale/Sarasota/Ft. Myers. As the aircraft arrived in that vicinity and was preparing to turn to the east for the final to Miami to land to the east, the weather came across the airport and shut down the operation. As a result, the aircraft entered airborne holding and was given "expect further times from ATC" that continued into the future. Thus, the crew was faced with an indefinite situation as to when they would be released to proceed into Miami.

It was not until this point that the Captain contacted the aircraft's dispatcher and advised the dispatcher that the reroute the Captain and the dispatcher had agreed upon had been refused by an ATC sector, that the aircraft had ended up back on its original filed route of flight, and that they had encountered airborne holding. The dispatcher's attention had been diverted to another situation and he had not noted the ATC-initiated reroute. Thus, at that point the aircraft was holding with thunderstorms between its position and the intended destination.

What complicated this scenario was that Sarasota, Ft. Myers, Ft. Lauderdale and West Palm Beach, which were all of the other usable alternate airports for this aircraft, were either unusable due to thunderstorms or were now north of the weather as well. The aircraft was basically trapped south of its intended destination and south of its usable alternates. (This aircraft was not authorized to use the Key West airport.) Consequently, the crew was faced with a situation of being very low on fuel with limited options in terms of available diversion airports. The aircraft finally broke through the line of thunderstorms as the weather passed south of Miami, and was able to land at Miami. However, they picked up significant turbulence going through the line of weather, producing a very uncomfortable ride for the crew or the passengers as the aircraft passed through severe turbulence.

It is also important to understand that the dispatcher working this particular flight on this day had about 30 other flights that he was responsible for at that time, and felt as though this situation had been resolved and had turned his attention to other situations that required his attention.

Important Features Illustrated by the Scenario | This scenario provides an example of one of the ways in which the air traffic management system has been decomposed into subtasks to reduce the cognitive complexity for individuals. This particular scenario also illustrates one potential weakness of such a decomposition, the reliance on individuals to appropriately decide when there is a need for interaction (i.e., when the decomposition is inadequate).

One response to such an incident would be to attempt to improve judgments about when to interact by improved training or more clearly defined procedures. A second would be to maintain the existing task decomposition, but to give everyone better access to critical elements of the bigger picture (such as weather), so that they could better judge when there is a need to interact with the other system operators. Another would be to develop technological support tools such as an "intelligent" alerting system that informed the dispatcher when a flight has begun to deviate "significantly" from its original route. A fourth would be to try to integrate decision making, abandoning or partially abandoning the task decomposition strategy. All of these approaches have strengths and weaknesses, and merit serious consideration for this specific scenario. The remainder of this paper, however, will focus on the fourth approach, and will do so in another ATM context, concerned with pre-flight planning and traffic flow management.

Design of the Air Traffic Management System | Historically, traffic flow management (TFM) has primarily a function under the control of the FAA, with traffic managers at various facilities making decisions about what routes could be flown by the flights scheduled by the airlines (Odoni, 1987). In recent years, however, there has been an emphasis on giving the airlines greater flexibility, based on the assumption that the airlines have better information about the costs of alternative flight plans and should therefore be in a position to make better decisions about the economics of alternative flight plans. This shift in essence changes the task decomposition as, under such changes, airline dispatchers must consider a much larger set of factors if they are to in fact improve performance. Issues surrounding such a shift are discussed below in terms of alternative system architectures for accomplishing it.

Alternative System Architectures | Alternative architectures for the air traffic management (ATM) system that change the decomposition of tasks for flight planning can be grouped into three categories (Smith, et al., 1997):

•   Management by directive (where FAA traffic managers simply inform an airline regarding the route that can be flown by a particular flight);

•   Management by permission (where there is a default flight plan assigned by the FAA, which can be revised if the airline operations center requests an alternative and receives permission from FAA traffic management staff);

•  Management by exception (where the airline operations center can simply file the flight plan that it desires for a given flight).

Over the past several years, the ATM system has been evolving from a system where management by directive was the predominant form of interaction to a hybrid system including examples of all three forms of interactions.

Control by Permission | The first major change arose in 1992, with a shift in management by directive to management by permission. Specifically, FAA Advisory Circular 90-91 established a formal procedure allowing the airlines to request non-preferred routes (routes for flights that differed from the FAA assigned preferred routes). Under this procedure, an airline could send a message via teletype to the FAA's Air Traffic Control Systems Command Center (ATCSCC) requesting an alternative route for a particular flight. A specialist at ATCSCC would then evaluate this request, checking with traffic managers at the involved regional enroute air traffic centers and, based on their input, would approve or disapprove the request.

This shift to management by permission gave the airlines a means for improving efficiency, because they had better information for determining the most economical flight plans for their aircraft. It still left the locus of control with the FAA traffic managers, however, as they had to individually approve all requested alternative routes. These approvals were made based on considerations of safety (avoiding excessive traffic bottlenecks) and overall efficiency in traffic flows. Thus, this shift left the basic task decomposition the same, but provided a procedure for increasing the frequency of interactions between traffic managers and dispatchers.

This new paradigm was viewed very positively by both the airlines and the FAA. One airline, for example, reported that in one year, it submitted 15,279 requests for non-preferred routes and that 75% of these requests were approved. These approvals resulted in an estimated savings of 13,396,510 lbs. of fuel. Studies by Smith, et al. identified a number of factors that appeared to contribute to this success.

Factors Contributing to Success | The first factor concerned matching the locus of control with access to relevant information. The criticism of prior procedures was that, under the management by directive paradigm, FAA traffic managers were making decisions that did not take into concern the airlines' business concerns. Thus the claim was that, for any given flight, there could be a number of equally acceptable flight plans from the perspective of safety and overall system efficiency, and in such cases the FAA was making the choice without the benefit of any input from the airline about its economic considerations. By leaving the ultimate decision up to FAA traffic managers, who had information and experience regarding potential traffic bottlenecks, but by allowing the airlines to indicate their preferences based on economic concerns, safety was ensured while economics were improved.

As with any system architecture, however, supporting arguments based on high-level considerations do not, by themselves, ensure that the architecture will be successful. The details of its implementation are equally important. Four major factors appeared to contribute to the success of this program:

•   Implementation of communication channels that led to the development of a shared understanding of goals, problems, constraints and solutions;

•   Distribution of responsibilities to a number of different individuals;

•   Incorporation of feedback and process control loops;

•  Design of positions with new roles and responsibilities.

In terms of the earlier discussion regarding task decomposition, what this did was to maintain most of the decomposition, in the sense that both the FAA traffic managers and airline dispatchers still had to analyze alternative routes from their own perspectives. The routine interactions, however, gave both groups a broader understanding of the factors considered by the other group, resulting in more effective and efficient interactions when they were likely to be productive (i.e., when the task decomposition was inadequate and there was a need for interactions between both groups in order to determine the best solution).

Limitations of Control by Permission | The primary weakness of this paradigm was that it was manpower intensive (requiring extra staffing to support the additional interactions), and was thought by the airlines to at times be excessively conservative in terms of the approval of requests for alternative routes. As a result, the system evolved further in 1995 to give the airlines further flexibility using a different "architecture".

Control by Exception | Although the use of the "control by permission" architecture was viewed as a significant improvement, its perceived limitations were sufficient to result in a followup program based on "control by exception". This new program, known as the expanded National Route Program, allowed the airlines, subject to certain constraints, to simply file the routes that they preferred for particular flights. FAA traffic managers would then monitor conditions, watching for conditions (such as severe weather) where the program had to be canceled temporarily for particular portions of the country. Tactical changes could also be initiated after the flight was enroute by FAA air traffic controllers (as well as by airline pilots and dispatchers with the concurrence of the responsible air traffic controllers). Unlike the earlier shift to "control by permission", this architectural change therefore significantly altered the historical task decomposition, requiring airline dispatchers to now consider factors (such as the prediction of air traffic bottlenecks) that in the past had been handled largely by FAA traffic managers.

To evaluate the impact of this architectural change, two studies were conducted dealing with the impact of the expanded National Route Program (NRP) on fuel consumption. The motivation for this study came from two sources. First, dispatchers at a number of airlines, as well as traffic managers at enroute air traffic control centers (ARTCCs), provided numerous examples of how flights filed under the NRP were sometimes given significant amendments, and suggested that some of these changes occurred on a regular basis. In at least some cases, the changes were clearly initiated by the ATC system to deal with traffic congestion. Along these lines, dispatchers made comments like:

"Under the expanded NRP, it's like shooting ducks in the dark."
"The problem with the expanded NRP is that there's no feedback. Nobody's getting smarter. Someone has to be responsible for identifying and communicating constraints and bottlenecks."
"It used to be the weather that was the biggest source of uncertainty. Now it's the air traffic system."

In short, the dispatchers appeared to be indicating that the shift in their tasks gave them more flexibility, but did not give them the information and tools necessary to integrate considerations of air traffic (one of the major factors that used to be handled primarily by the FAA traffic managers) into their decision making.

As a specific example, one dispatcher indicated that NRP flights from Washington National to Cincinnati frequently have a problem because of the strategy used by ATC to deal with crossing traffic: "It happens to us all the time. We file the flights at 35 or 39 [altitudes of 35,000 or 39,000 feet] and they're held at 23, 25 and 27. They don't tell us ahead of time that it's going to happen."

A second example of how traffic bottlenecks can affect NRP flights was provided by a traffic manager: "Quite often ... 8-10 extra aircraft are on this northern route to DFW [from Southern California to Dallas flying north of White Sands into the northwest cornerpost at the Dallas-Fort Worth airport] during the noon arrival rush [noon local time]. This causes a sector saturation problem in ZFW Sectors 93 and 47 [two Dallas-Fort Worth (DFW) air traffic control sectors]. To relieve this volume problem, the ZFW TMU [Traffic Management Unit] moves 5 aircraft back to the south route [south of White Sands] via CME.TQA.AQN.DFW [a sequence of navigational fixes into the southwest cornerpost of the Dallas-Fort Worth airport]. This longer route of flight, plus the fact that DFW is in a south flow (meaning these flights will spend more time flying below 10,000 feet), will reduce fuel savings or negate them all together for this bank of flights."

Thus, anecdotal evidence suggested that traffic bottlenecks were arising that impacted the efficiency of NRP flights, raising questions about the effectiveness of this new decomposition of tasks. To gain further insights into this concern, two followup studies were conducted. These are described below.

Study 1: Analysis of Predicted vs. Actual Fuel Consumption | To look for evidence of such inefficiencies, we collected data from a major airline on all of their flights filed over a 5 month period. These data were used to compare predicted fuel consumption on NRP routes with predicted fuel consumption on FAA preferred routes and also with actual fuel consumption. In the following discussion, a "flight" is defined to be a particular combination of an origin, destination, Ptime (scheduled departure time) and equipment type. Thus a given flight could have a new instance filed each day. Predicted and actual fuel consumptions were from wheels-up to wheels-down.

Predicted fuel consumptions were first analyzed, comparing performances on FAA preferred routes with the filed NRP routes. 21,334 flight instances were filed by this airline under the NRP during this time period. The average predicted fuel savings per day during this time period ranged from 2.3% to 6.0%. The total predicted savings was 17,723,329 lbs.

Comparison of Predicted vs. Actual Fuel Consumption | Given the anecdotal evidence outlined earlier, however, it seems possible that these predictions overestimate actual fuel savings for some flights, since the computer's predictions do not take into account the new reroutings that might occur as a result of filing an NRP route and then encountering a traffic bottleneck while enroute. Consequently, we also compared predicted with actual fuel consumption.

To ensure adequate statistical power, only flights with at least 20 instances were considered. There were 267 such flights. A statistical analysis indicated that 94, or 35%, of these flights routinely burned more fuel than predicted (P<0.05). Of these 94, 21% routinely burned more extra fuel than was supposed to be saved by flying the NRP route instead of the FAA preferred route. (The predictions generated 75 minutes prior to scheduled departure were used for this comparison.) The flight from DFW to SNA at 1645 UTC (flying an MD80), for instance, on average burned 1013 lbs. of fuel more than predicted. As a result that flight, which on average was supposed to save 759 lbs. of fuel compared to the FAA preferred route (a predicted 4% savings), actually burned 254 lbs. more than the prediction for the FAA preferred route (a 1.3% loss).

These data also indicated that the city pair that most often had flights with regular problems was LAX to DFW. Seventeen of those flights routinely burned more fuel that predicted.

Study 1. Implications | Minimally, these data indicate that there was some sort of a problem associated with 35% of the flights filed by this airline under the NRP during this time period. One possibility would be an underlying inaccuracy in the prediction model for one or more of these flights, over and above any new problems introduced by use of the expanded NRP. If, however, we assume that the prediction model provides unbiased estimates (after discounting any new problems introduced by use of the expanded NRP), then these data indicate that the actual benefits in terms of fuel consumption from the use of the NRP are less than predicted by this airline.

Study 2: A Detailed Observational Study of LAX-DFW Flights | As mentioned above, the city pair that most often encountered problems was LAX-DFW. We therefore decided to study it in detail in order to collect more detailed data on the nature of the problems with NRP flights for this city pair, and to better quantify the impact of these problems.

Methods | Four students from the Aviation Department at Ohio University collected data from June 22, 1996 to August 23, 1996 on the performances of flights from LAX-DFW. Flights with five different scheduled departure times (Ptimes) were studied (1400, 1415, 1445, 1515, and 1810 Universal Coordinated Time or UTC). The students collected data on predicted and actual fuel consumptions and observed each flight instance on the ASD to record any flight amendments. They also interviewed pilots immediately after each flight to help document the causes of any flight amendments that arose, and to assess the pilots' understanding of the NRP.

Results | The resultant observations quickly made it clear that the underlying problem was the rerouting described earlier. Very briefly, what happens is:

•   A flight instance is filed under the expanded NRP along a route north of White Sands (special use airspace) to the northwest cornerpost at DFW;

•   While that flight is enroute, the Air Traffic Management (ATM) system decides that there is likely to be a sector saturation problem in the Turkey or Falls high sectors when the flight reaches that point as it approaches the northwest cornerpost into DFW;

•   To deal with that problem, the flight or flights with the most southerly routes that are flying to the northwest cornerpost are rerouted south of White Sands to the FAA preferred route so that they will approach DFW via the southwest cornerpost.

The discussion below provides details on this problem.

Table 1. Percentage of Flights Flying the FAA Preferred Route (Pref Route) and NRP Routes with or without Cornerpost Swaps. (Ptime is Universal Coordinated Time or UTC) |

Table 2. Expected vs. Actual Fuel Savings for Those Flights Filed Under the NRP that were Rerouted from the Northwest to the Southwest Cornerpost. (Ptime is UTC; Savings are the % reduction or increase relative to the predicted fuel consumption for the FAA preferred route that day) |

Cornerpost Swapping. Table 1 indicates the frequency with which the cornerpost swap occurred for the different flights that we observed. (Keep in mind that this swap usually occurs before White Sands, not as the flights are approaching the airport.) The results indicate that the flights that arrive at DFW for the noon rush (flights that are arriving into DFW around noon local time, and that have scheduled departure times or Ptimes of 1400 and 1415 UTC) are particularly affected. 33-39% of the flights during that time period fell into that category and were rerouted south of White Sands to the FAA preferred route. Table 2 indicates the impact of this rerouting on overall savings for the NRP flights filed at particular Ptimes for those instances where an NRP flight was actually rerouted south of White Sands. All of these flights on average burned more fuel than was predicted if they had been filed on the FAA preferred route. On average, for example, it cost an additional 1502 lbs. of fuel each time the flight at 1400 UTC was rerouted to the southwest cornerpost. A statistical test comparing actual with predicted fuels consumptions for these flights was significant (p<.05) for the Ptimes of 1400, 1445 and 1810.

Conclusion | The case studies reviewed in this paper serve to illustrate several points:

•  One classic strategy for reducing cognitive complexity in the ATM system has been to decompose the system into subtasks, and to assign these tasks to different individuals. Then, in those circumstances where the assumption of independence among these subtasks breaks down, it is necessary for the responsible individuals to interact with each other;

•   A drawback of such a decomposition strategy is that the responsible individuals may not recognize the need for such interaction. This can result in problems from either a safety or efficiency standpoint, as illustrated by the example of the flight from Dallas to Miami;

•   A second drawback is that, because the assumption of independence made during the task decomposition is at best an approximation, and because the "as required" interactions of individuals to deal with inadequacies in this task decomposition typically only partially compensate for the approximations, although overall system performance may be good, it is not likely to achieve its theoretical "optimal";

•   Because of these drawbacks, a variety of alternative architectures or task decompositions are now being explored within the ATM system. Two such architectures, "control by permission" and "control by exception" are illustrated in the context of preflight planning. The first architecture, "control by permission", attempts to improve performance by maintaining the traditional task decomposition, but by improving the interactions between traffic managers and dispatchers to cope with the limitations of this decomposition. The second architecture, "control by exception", represents a major change in task decompositions;

•   Studies of use of the second architecture, "control by exception", provide cautions about the need to fully consider the impact of alternative task decompositions on information requirements and on the cognitive complexity of the newly defined tasks. These studies caution that, without such considerations, the expected move to an "optimum" level of performance may in fact not be fully achieved.

Acknowledgments | This work was supported by the FAA Office of the Chief Scientist and Technical Advisor for Human Factors (AAR-100) and NASA Ames Research Center. We would like to express special appreciation to Larry Cole, Eleana Edens, Tom McCloy, Mark Hoffman, Roger Beatty, Joe Bertapelle, Rob Blume, Scott Ridge, Moira Hoban Edwards and John Tittle.

References | Odoni, A.R. (1987) The flow management problem in air traffic control. In A.R. Odoni, L. Bianco, and Szego (eds.), Flow Control of Congested Networks. Berlin: Springer-Verlag.

Smith, P.J., McCoy, C.E., Orasanu, J., Billings, C., Denning, R., Rodvold, M. Gee, T. and Van Horn, A. (1997) Control by permission: A case study of cooperative problem solving in the interactions of airline dispatchers with ATCSCC. Air Traffic Control Quarterly, 4, 229-247.