From trees to forest: relational complexity network and workload of air traffic controllers

Abstract

In this paper, we propose a relational complexity (RC) network framework based on RC metric and network theory to model controllers' workload in conflict detection and resolution. We suggest that, at the sector level, air traffic showing a centralised network pattern can provide cognitive benefits in visual search and resolution decision which will in turn result in lower workload. We found that the network centralisation index can account for more variance in predicting perceived workload and task completion time in both a static conflict detection task (Study 1) and a dynamic one (Study 2) in addition to other aircraft-level and pair-level factors. This finding suggests that linear combination of aircraft-level or dyad-level information may not be adequate and the global-pattern-based index is necessary. Theoretical and practical implications of using this framework to improve future workload modelling and management are discussed.

Abstract

Practitioner Summary: We propose a RC network framework to model the workload of air traffic controllers. The effect of network centralisation was examined in both a static conflict detection task and a dynamic one. Network centralisation was predictive of perceived workload and task completion time over and above other control variables.

Keywords::

• relational complexity network
• air traffic control
• workload
• network centralisation

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

^1. Email: [email protected].

^2. We used the binary coding instead of the 0–4 coding system used by of Boag et al. (2006) for the following reasons. First, we believe that it is mathematically feasible to use the ordinal metric and a weighted network approach. However, as more different patterns (such as degree distribution) may occur in a weighted network, it requires more varied scenarios to offer empirical evidence which can be very difficult for researchers to design and also burdensome for participants. It might be more practical to use such a framework if certain ‘big data’ from the real ATC operation centres can be obtained. Second, most of the RC>0 conditions in our scenarios (94%) were actually the RC = 2 conditions according to the original Boag et al. (2006) metric. In this regard, we believe that this dichotomous coding would not reduce much information.

^3. These filler scenarios were used for testing a different hypothesis; however, including them in the current analysis does not change the major findings. Interested readers may contact the authors for further information about the analysis including these scenarios.

^4. This is the ratio between pairs still in conflict after they provided resolution and pairs in conflict at the start of scenarios. The conflict detection rate (pairs reported to be in conflict using the detection function), on the other hand, was 40.4% for Study 1 and 49.8% for Study 2. They were much higher than the three-aircraft conditions in Boag et al. (2006) (about 10%) using experienced controllers but similar to the findings in an earlier study (about 40.0%) (Landry, Sheridan, and Yufik 2001) using amateurs and more aircraft. It must be mentioned that our tasks were relatively more difficult since: (1) nearly all conflicting pairs involved both horizontal conflict and altitude traversing; (2) the conflict occurrence rate was quite low (8%), and (3) the participants were not able to see their detection results, therefore they may have forgotten the data that had already been entered into the system. More importantly, in many of our scenarios, one aircraft can cause many potential conflicts, reporting one of them can be considered by participants to be enough as resolving it can prevent all other related conflicts. This is especially the case for Study 2 in which early intervention can prevent some pre-arranged conflicts to appear at all. If we adopt such a lenient criterion by treating detecting one pair among all related pairs (sharing one airplane in common) as a hit for them all, then the detection miss rate will sharply drop to 22.6% and 38.3% for Studies 1 and 2, respectively. If we further consider resolution, the rate will drop to 14.6% and 21.4%, respectively. None of them is correlated with completion time.

Additional information

Funding

This research was supported by the National Basic Research Program of China [grant number: 2011CB302201], the Key Research Program of the Chinese Academy of Sciences [grant number: KJZD-EW-L04] and the Natural Scientific Foundation of China [grant number:31300875 and 31070915].