Predictive Sectorization and Bayesian Optimized Consensus for Admission Control in Autonomous Airspace Operations Aditya Dhodapkar∗ , Avery Smidt† , Aaron Verkleeren∗ , Stacy Patterson∗ , Carlos A. Varela∗ ∗ Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY, USA † Quaternion Consulting Inc., Herndon, VA, USA
arXiv:2604.17063v1 [cs.DC] 18 Apr 2026
{dhodaa, verkla}@rpi.edu, [email protected], {sep, cvarela}@cs.rpi.edu Abstract—Conventional air traffic control divides airspace into specific regions, creating a scaling bottleneck as traffic grows. Choosing how to partition airspace is not straightforward because grid size affects workload, handoff frequency, and the capacity of whatever coordination mechanism operates within each sector. We present a three stage pipeline that automates sectorization and sector coordination while preserving human oversight. First, a two stage XGBoost classifier predicts the optimal 3D grid configuration from 23 location-agnostic traffic features, achieving 91.38% accuracy on a 65,000 sample dataset derived from Federal Aviation Administration System Wide Information Management replays. Second, a leaderless Paxos consensus protocol lets aircraft coordinate sector entries among themselves, maintaining above 96% entry success with low near mid-air collision rates across all tested configurations. Third, Bayesian Optimization with a Gaussian Process surrogate tunes eight protocol parameters per airport in 50 trials, revealing that each traffic environment requires a qualitatively different configuration. The resulting pipeline offers a practical path toward scalable, autonomous airspace management as traffic demand outpaces controller capacity.
I. I NTRODUCTION Air traffic control operates reliably almost all the time. The problem is what happens when it does not. The Federal Aviation Administration (FAA) has reported staffing shortages at major facilities for years [1], [2]. Controllers work overtime and six day weeks to maintain coverage. In Europe, industrial action has repeatedly produced network level delays and cancellations [3]. Traffic surges and weather induced rerouting compound controller workload simultaneously, increasing handoff frequency, reducing separation margins, and shortening available decision time. In the air traffic control system, one controller is responsible for every aircraft in their sector. A controller typically manages 8–12 aircraft simultaneously, a manageable cognitive load, but peak periods can push this above 18, producing an overwhelming workload. Research on dynamic density [4] and cognitive complexity [5] has shown that workload depends not just on aircraft count but on proximity, closure rates, heading changes, and crossing streams. Dynamic Airspace Configuration (DAC) has been proposed as a remedy [6], adjusting sector boundaries in response to evolving demand. But existing DAC approaches rely on computationally expensive graph cut or clustering
optimizations [7] that are difficult to run at the cadence required for tactical decisions. Sectorization research spans algorithmic boundary placement [7] and demand responsive configuration [6], both relying on workload metrics rooted in dynamic density [4] and cognitive complexity [5]. For decentralized coordination, Paul et al. introduced conflict aware flight planning [8], [9] with formal verification of safety under asynchronous networks [10], [11]. We build on this work in our Decentralized Air Traffic Control (DATC) protocol. Bayesian Optimization has been applied to expensive black box tuning across domains [12], [13], but not to consensus protocol parameters in aviation. We treat sectorization as a supervised classification problem and pair it with a decentralized consensus protocol, keeping controllers in charge of policy while offloading routine separation tasks. Our proposed pipeline (Fig. 1) has three stages. Stage 1: A two stage XGBoost predictor takes as input 23 aggregate traffic features derived from a snapshot of the airspace (density, proximity, flow direction, altitude mix, etc.) and maps them to the optimal grid configuration, i.e., the number of rows and columns of a uniform rectangular partition that divides the airspace into sectors. Stage 2: The predicted grid defines sector boundaries for a leaderless Paxos consensus protocol in which aircraft already occupying a sector coordinate with an arriving aircraft to decide whether its proposed flight plan can be admitted without causing conflicts. Stage 3: Bayesian Optimization (BO) with a Gaussian Process surrogate tunes eight protocol parameters per airport to maximize admission success while minimizing holding patterns, speed modifications, and retries. To our knowledge, this is the first work to combine learned sectorization, decentralized consensus, and automatic protocol tuning into a single end to end pipeline. SWIM Data
XGBoost Predictor
Consensus Protocol
Bayesian Optimization
Fig. 1: Pipeline overview. Our contributions are: 1) A two stage XGBoost classifier achieving 91.38% accuracy on 25 class sectorization prediction using 23
TABLE I: Feature summary (23 features).
location-agnostic features, outperforming Random Forest, LightGBM, CatBoost, and single stage XGBoost baselines on the same dataset. 2) A DATC consensus protocol contribution: a) A standalone integration of three protocols (Discovery, Synod, TAP) [9], [14] into the DATC consensus framework with reliability extensions. b) Validation on both synthetic and real JFK System Wide Information Management (SWIM) traffic across grid sizes from 2×2 to 16×16 and up to 160 aircraft. 3) A Bayesian Optimization framework that tunes eight protocol parameters per airport in 50 trials, demonstrating that no single default configuration is sufficient across traffic environments. A full treatment of each stage, including additional baselines, ablations, and extended evaluation, is available in the first author’s master’s thesis [15]. The remainder of this paper is organized as follows. Section II presents the sectorization predictor, Section III describes the consensus protocol, and Section IV covers the Bayesian Optimization framework. Section V discusses findings and limitations, and Section VI concludes.
Range
Aircraft count Mean pairwise dist. Altitude std. dev. Mean min. sep. Mean | sin ψ| Circ. std. dev. Hour Region radius
[0, 150] [0, 80] NM [0, 15k] ft [0, 60] NM [0, 1] [0, 1.5] [0, 23] 100 NM
Engineered (9) congestion index traffic alt complexity hotspot indicator density sq log proximity traffic level proximity level risk normalized flow direction
density/proximity density × alt mix risk/proximity density2 ln(1+prox) Binned density Binned proximity risk/density Axial alignment
– – – – – {0, 1, 2} {0, 1, 2} – [0, 1]
C. Model Architecture The model uses a two stage XGBoost [18] architecture. Stage 1 is a binary classifier (500 trees, depth 6, learning rate 0.05) that separates 1×1 from non-1×1 samples with 99.29% recall. This near perfect recall ensures the model almost never predicts single sector when multiple sectors are needed, which would cause controller overload. Stage 2 is a 24 class classifier (800 trees, depth 14, learning rate 0.02, L1 = 0.1) that predicts the exact configuration. At inference, a sample classified as 1×1 by Stage 1 receives (R, C) = (1, 1) directly; otherwise Stage 2 predicts the full configuration. Because 1×1 dominates raw SWIM traffic, a single stage 25 class model learns to default to the trivial case and achieves only 85.12% accuracy. The two stage design filters these out first so Stage 2 only discriminates among configurations that actually need partitioning.
A. Problem and Data Given a traffic snapshot over a 100 × 100 Nautical Mile (NM) study region, we predict the optimal horizontal grid (R∗ , C ∗ ) from {1, . . . , 5}2 (25 configurations) with three fixed altitude layers (0–9,999 ft, 10,000–17,999 ft, 18,000+ ft). The 100 NM side length ensures that even the finest predictor grid (5×5) produces 20 NM sectors, large enough for adequate dwell time at cruise speeds. Labels are generated by exhaustive grid search over a composite objective combining occupancy variance, handoff counts, and risk pair penalties: (R,C)
Description
Raw (8) traffic density avg proximity altitude mix conflict risk primary flow dir flow concentration time of day airspace size
Time/day (6) day weekday, day weekend, time {morn, noon, eve, night}
II. S ECTORIZATION P REDICTOR
(R∗ , C ∗ ) = arg min J(R, C, L=3).
Feature
(1)
D. Results
Traffic was recorded from 100 × 100 NM regions centered on five major US airports (JFK, ATL, ORD, DFW, LAX) and five en route cruising corridors with consistently high traffic flow, via the FAA SWIM program [16]. In most raw snapshots, traffic density is low enough that 1×1 (no sectorization) dominates the unbalanced dataset. We augment underrepresented configurations with the Synthetic Minority Over-sampling Technique (SMOTE) [17], producing approximately 65,000 samples (40,000 simulation, 25,000 synthetic) split 70/15/15 for training, validation, and testing. Most configurations receive approximately 2,700 samples each, with 1×1 at 5,350 and 5×5 at 300.
The two stage model achieves 91.38% overall accuracy (91.1±0.3% under five fold cross validation). Stage 2 predicts rows correctly 95.5% of the time and columns 94.8%, but both must be correct simultaneously, producing the 90.91% combined non-1×1 accuracy. Table II compares against four alternatives trained on the same data with the same split. TABLE II: Model comparison on the held out test set. Model Single Stage XGBoost Random Forest LightGBM CatBoost Two Stage XGBoost
B. Features We extract 23 location-agnostic features organized into three groups (Table I). No geographic identifiers are included, forcing the model to learn from traffic characteristics alone.
Accuracy 85.12% 90.17% 90.69% 90.83% 91.38%
Confusion matrix analysis reveals three systematic error patterns: (1) low partition confusion (1×2 ↔ 2×1, both
2
represent a single bisection differing only in orientation), (2) row count adjacency (e.g., 5×1 → 4×1 when density profiles overlap), and (3) column swap (3×4 ↔ 3×5, correct row with off by one column). In practice, these neighboring configurations produce similar workload distributions, limiting the operational impact of such errors. The full 25-class confusion matrix is available in the thesis. The top three features (traffic level, traffic density, traffic density squared) account for over 79% of total gain, confirming that sectorization is primarily driven by traffic volume. Inference requires less than 1 ms per prediction on a single CPU core.
adjusts ground speeds. If speed adjustment fails, a holding pattern is assigned. If all alternatives are exhausted, the entry is denied. When multiple aircraft propose simultaneously, they can repeatedly preempt each other by incrementing proposal numbers in an unbroken cycle; we mitigate this livelock risk with randomized exponential backoff drawn from [nackBackoffMin, nackBackoffMax]. Phase C (TAP, Two Phase Acknowledge Protocol): The admitted plan is reliably disseminated to all sector members in two rounds: LEARN/LEARNT confirms every member received the decided value, then AK/ACK confirms every member knows that every other member received it, ensuring every aircraft holds an identical view of the admitted plan set before the next admission. Safety and liveness. Safety here means that no two conflicting flight plans are ever both admitted, even under arbitrary message delays, reordering, and loss. This is guaranteed by Paxos quorum intersection: any two majority quorums share at least one member, so at least one voter in any new round has seen the latest admitted plan [10], [19]. TAP ensures all sector members share an identical view of the admitted set, preventing divergent views. Our primary results address liveness: eventual consensus progress via livelock prevention [8] and timely TAP completion under probabilistic message latency [9]. Exit protocols. When an aircraft leaves a sector, it runs the second two phases (Synod and TAP) with an exit flag, proposing removal of its plan from the admitted set rather than addition. Discovery is unnecessary because the exiting aircraft is already a sector member. The Synod round agrees on the removal, and TAP disseminates the updated state. A denied entry aircraft that has exhausted all alternates initiates an emergency exit (C3 exit, abandoning its expected arrival slot) to abort the attempt, remove itself from the sector’s tracking sets, and notify all occupants that the entry has been abandoned. This symmetric design means entry and exit use identical protocol machinery, simplifying both verification and implementation.
III. C ONSENSUS P ROTOCOL A. Background and Architecture The predicted grid defines sector boundaries. Within each sector, aircraft coordinate entries using the DATC protocol [9], [14], which adapts Lamport’s Paxos [19] into a three phase admission procedure. Paul et al. developed conflict aware flight planning with formally verified correctness guarantees [8], [11] and proved eventual and timely consensus under specific network constraints regarding message delays, reordering, and loss [10]. We integrated the three constituent protocols (Discovery, Synod, TAP) into a standalone C++ library driven by a discrete event simulator with a per aircraft engine architecture. Each aircraft runs its own consensus engine containing proposal state, message history, and deduplication sets. The engines communicate through a shared event queue ordered by (time, aircraft ID, event type) tuples with integer picosecond timestamps for deterministic execution. Reliability extensions include fuel tracking with diversion logic and randomized exponential backoff for livelock prevention; full implementation details and the correctness argument appear in the thesis. B. Three Phase Protocol Every sector admission runs through three sequential phases. Phase A (Discovery): The entering aircraft broadcasts INIT REQ to discover current sector occupants. REQ ACK responses establish the quorum (majority of occupants needed to agree, ensuring safety under message loss) and provide each occupant’s current flight plan. The proposer uses this information to compute a provably collision-free flight plan before entering Synod. Phase B (Synod): A single round of the Synod protocol [19] determines whether the proposed plan can be admitted. The proposer sends PREPARE with a unique proposal number; occupants respond with PROMISE, including the number and value of the highest-numbered proposal they have already accepted (if any). Once a quorum promises, the proposer sends ACCEPT: if any PROMISE carried an accepted value, the proposer must re-propose the value with the highest proposal number; otherwise it proposes its own collision-free plan with a certificate of collision-freedom. If no conflict-free plan can be constructed, a formally verified backtracking algorithm [8]
C. Evaluation We evaluate the protocol through parametric sweeps across grid configurations and traffic counts on both synthetic trajectories (grid sizes 2×2, 4×4, 8×8, 16×16; traffic counts 10 to 160) and real JFK SWIM data (grid sizes 2×2 through 6×6; traffic counts 10 to 80). Note that these sweep ranges are broader than the predictor’s {1, . . . , 5}2 space; the larger grids stress test the protocol beyond predicted configurations. All synthetic configurations maintain above 96% entry success through 160 aircraft, with the 8×8 grid holding 100% across the entire range (zero denied entries). Fig. 2 shows the conflict resolution breakdown for the JFK sweep. Holding patterns are the most frequently used mechanism because speed modification, though attempted first, often fails to resolve geometric conflicts. Denied entries remain a small fraction (≤ 3 across all JFK configurations; the synthetic sweep shows a similarly small fraction of ≤ 12),
3
Fig. 2: Conflict resolution breakdown for JFK real world traffic. Stacked bars show holdings, speed modifications, and denied entries.
Fig. 4: Wall clock time vs. aircraft count (JFK). Horizontal bars show min to max range across grid sizes. The 5×5 outlier at 80 aircraft (1,752 s) is caused by Paxos livelock in a single sector.
Fig. 3: NMACs vs. aircraft count for JFK real world traffic. The 2×2 grid produces the fewest NMACs because most converging approach traffic stays within a single sector. Fig. 5: Intra-sector consensus (Phase 1) vs. inter-sector coordination (Phase 2) time as a function of grid size (JFK, 60 aircraft). Phase 2 dominates and grows with finer grids; the secondary axis shows the corresponding decrease in maximum quorum size.
confirming that the holding plus speed mechanism resolves the vast majority of conflicts. Fig. 3 shows Near Mid-Air Collision (NMAC) counts for the JFK sweep: the 2×2 grid produces the fewest NMACs (0– 1 across all counts) because most converging approach traffic stays within a single sector and is handled by one consensus round, while finer grids split approach corridors across sector boundaries, creating inter sector separation violations. This relationship reverses in the synthetic sweep, where coarser grids produce the most NMACs (40 at 160 aircraft for 2×2, down to 8 for 16×16) because fewer internal boundaries means fewer opportunities for the protocol to detect converging traffic. The reversal highlights that the optimal grid size for safety depends on traffic geometry, not just density, motivating the use of the XGBoost predictor to adapt the grid to actual traffic patterns. Fig. 4 shows wall clock execution time for the JFK sweep as a Gantt style range chart. Each horizontal bar spans the min to max time across grid sizes at a given aircraft count, with diamond markers for individual grids and a vertical line at the median. Times scale super linearly with aircraft count, driven by quadratic growth in pairwise conflict checks. Most configurations at 80 aircraft complete in under 200 s. The striking outlier is the 5×5 grid at 80 aircraft (∼1,752 s), caused by Paxos livelock in a single high density sector. This occurred once across all experiments, confirming it as a tail risk rather
than a systemic limitation. Fig. 5 decomposes execution time into its two phases to expose the fundamental sectorization tradeoff. Phase 1 (intrasector consensus) remains under 4 s across all grid sizes because finer grids reduce quorum sizes, making each Paxos round faster. Phase 2 (inter-sector coordination) dominates and grows monotonically: more sectors mean more boundary crossings and handoff negotiations, rising from 31.6 s at 2×2 to 66.7 s at 6×6. Total wall clock time is therefore minimized at the coarsest grid (2×2, 35.0 s), but as shown in Fig. 3 this configuration incurs the highest NMAC rate. The 4×4 grid represents the practical sweet spot: it keeps total time within 43% of the minimum while eliminating all NMACs at 60 aircraft. IV. BAYESIAN O PTIMIZATION A. Formulation The consensus protocol exposes eight tunable parameters: timeout duration, number of retry attempts, number of alternates, phase delay factor, NACK backoff minimum and maximum, Initial Request (IR) time, and solver timeout. Even
4
a coarse five level discretization of each parameter would require 58 = 390,625 simulations. Each evaluation requires a full 30 minute simulation, making gradient free black box optimization the appropriate regime [12]. The objective combines four terms: J(θ) = 100 · rsuccess − 30 · rhold − 10 · rspeed − 2 · n̄retry (2) where rsuccess is the admission throughput ratio, rhold and rspeed are the fractions of admissions requiring holding patterns or speed modifications, and n̄retry is the average retry count per initiated request. The coefficients were chosen based on operational reasoning: entry success is the dominant objective (×100), holding consumes more fuel than speed changes (×30 vs. ×10), and retries are invisible to end users (×2). Hard penalties of −100 for any NMAC and −50 for wall clock timeout override the continuous score. Derivations of the weight choices and per-airport sensitivity studies appear in the thesis. A Gaussian Process (GP) with a Matérn 5/2 Automatic Relevance Determination (ARD) kernel [20] models the objective surface. The ARD parameterization assigns a separate lengthscale ℓd to each input dimension; after fitting, dimensions with shorter lengthscales exhibit greater influence. Expected Improvement [13], [21] selects each trial. The budget is 50 trials per airport: 10 Latin Hypercube Sampling (LHS) [22] initial samples followed by 40 GP guided iterations. An initial campaign on ORD produced five timed out trials at parameter extremes. We tightened bounds to exclude the pathological region and enforce nackBackoffMin < nackBackoffMax.
Fig. 6: BO convergence trajectory for LAX (31 aircraft).
Fig. 7: BO convergence trajectory for DFW (60 aircraft). TABLE IV: Top three parameter importances per airport, derived from GP ARD lengthscales (normalized inverse lengthscale).
B. Results Table III summarizes optimal configurations across three airports with different traffic densities. TABLE III: Best BO configurations by airport.
Best Score Timeout Duration (s) IR Attempts Solve Timeout (s) Timeouts (/50)
LAX (31 acft)
ORD (47 acft)
DFW (60 acft)
189.8 0.50 10 10.0 0
127.1 0.50 15 1.0 5
100.8 2.93 4 9.1 1
Airport
Parameter
Importance
LAX (31 acft)
Solve Timeout NACK Backoff Min Start IR Time
36.7% 31.7% 11.4%
ORD (47 acft)
NACK Backoff Min Start IR Time Timeout Duration
64.3% 27.6% 3.1%
DFW (60 acft)
Timeout Duration NACK Backoff Min Alternates
45.6% 23.2% 14.8%
lengthscales. Solver timeout dominates at LAX (36.7%), while NACK backoff and timeout duration become more important at higher densities where contention drives retry behavior.
Figs. 6 and 7 show convergence trajectories for LAX and DFW. Across all three airports, LAX reaches its best score at trial 35, ORD at trial 40, and DFW at trial 19. The optimal configurations differ qualitatively: LAX favors many retry attempts with a long solver window, ORD requires a short solver timeout (1.0 s) to avoid pathological retry chains that caused 5 of its 50 trials to time out, and DFW favors few retries (4) with a longer timeout duration (2.93 s). This confirms that no single default setting works and that protocol tuning must be traffic density dependent. Table IV summarizes the top three parameter importances per airport, derived from the ARD kernel’s per parameter
V. D ISCUSSION Three specific findings stand out. First, the XGBoost predictor achieves 91.38% accuracy using only aggregate traffic features with no geographic identifiers, confirming that sectorization depends on traffic state rather than airspace identity and supporting deployment to unseen regions without retraining. Predicting from aggregate features also scales to larger configuration spaces where exhaustive scoring becomes impractical. Second, the NMAC analysis reveals that grid size interacts with traffic geometry in non-obvious ways. Uniform grids
5
that work well for random traffic can perform poorly on real approach corridors, and vice versa. This motivates the use of the predictor to select grid configurations adapted to actual traffic patterns rather than relying on fixed partitions. Third, the livelock outlier at 5×5/80 aircraft underscores why automatic parameter tuning is essential. BO identified NACK backoff as the most important parameter at ORD and DFW (64.3% and 23.2% importance), precisely the mechanism that prevents livelock. Manual tuning is unlikely to find these configurations given the eight-dimensional search space. Furthermore, Bayesian Optimization confirms that protocol sensitivity is environment dependent. The qualitative differences between optimal configurations at LAX, ORD, and DFW mean that manual parameter selection cannot scale. Automatic tuning is essential for deployment across diverse environments. Limitations. The simulation uses straight line flight paths and point mass dynamics, which limits realism near terminal areas where aircraft follow curved, altitude constrained procedures. However, the consensus protocol is trajectory agnostic: it operates on the sequence of waypoints and velocity vectors an aircraft submits, not on how those waypoints were produced. Replacing the straight line generator with a procedural or published flight track therefore requires no protocol changes. The same admission, conflict detection, and TAP dissemination logic applies to any flight plan representation. Communication is deterministic, reliable, has no faulty processes, and no latency jitter. The protocol does not currently handle aircraft emergencies, which would require priority override mechanisms. The BO budget of 50 trials per airport may miss fine grained optima. The objective function weights (100, 30, 10, 2) were set by intuition rather than derived from data.
agnostic design make the system a practical candidate for autonomous airspace management; codebase: [23]. R EFERENCES [1] U.S. Department of Transportation, Office of Inspector General, “FAA faces controller staffing challenges as air traffic operations return to prepandemic levels at critical facilities,” DOT OIG, Tech. Rep. AV2023035, Jun. 2023. [2] Federal Aviation Administration, “Air traffic controller workforce plan, FY 2025–FY 2028,” FAA, Tech. Rep., 2024. [3] EUROCONTROL, “Network operations report 2023,” EUROCONTROL, Tech. Rep., 2024. [4] I. V. Laudeman, S. G. Shelden, R. Branstrom, and C. R. Brasil, “Dynamic density: An air traffic management metric,” NASA Ames Research Center, NASA Tech. Memorandum NASA TM-1998-112226, 1998. [5] EUROCONTROL Experimental Centre, “Cognitive complexity in air traffic control: A literature review,” EUROCONTROL, Tech. Rep. EEC Note 04/03, 2003. [6] P. Kopardekar, K. Bilimoria, and B. Sridhar, “Initial concepts for dynamic airspace configuration,” in AIAA Aviation Technol., Integr., Oper. Conf. (ATIO), 2007. [7] P. Flener and J. Pearson, “Automatic airspace sectorisation: A survey,” Knowl. Eng. Rev., vol. 28, no. 3, pp. 293–314, 2013. [8] S. Paul, S. Patterson, and C. A. Varela, “Conflict-aware flight planning for avoiding near mid-air collisions,” in 38th AIAA/IEEE Digit. Avionics Syst. Conf. (DASC), San Diego, CA, USA, 2019. [9] ——, “Collaborative situational awareness for conflict-aware flight planning,” in 2020 AIAA/IEEE 39th Digit. Avionics Syst. Conf. (DASC), San Antonio, TX, USA, 2020, pp. 1–10. [10] S. Paul, G. A. Agha, S. Patterson, and C. A. Varela, “Eventual consensus in synod: Verification using a failure-aware actor model,” Innov. Syst. Softw. Eng., 2022. [11] S. Paul, C. McCarthy, S. Patterson, and C. Varela, “Formal verification of timely knowledge propagation in airborne networks,” Sci. Comput. Program., vol. 239, 2025. [12] B. Shahriari, K. Swersky, Z. Wang, R. P. Adams, and N. de Freitas, “Taking the human out of the loop: A review of Bayesian optimization,” Proc. IEEE, vol. 104, no. 1, pp. 148–175, 2016. [13] J. Snoek, H. Larochelle, and R. P. Adams, “Practical Bayesian optimization of machine learning algorithms,” in Adv. Neural Inf. Process. Syst., vol. 25, 2012. [14] S. Paul et al., “Formal verification of safety-critical aerospace systems,” IEEE Aerosp. Electron. Syst. Mag., vol. 38, no. 5, pp. 72–88, 2023. [15] A. Dhodapkar, “Predictive sectorization and Bayesian optimized consensus for admission control in autonomous airspace operations,” Master’s thesis, Rensselaer Polytechnic Institute, Troy, NY, USA, 2026, available: https://wcl.cs.rpi.edu/theses/dhodaa thesis.pdf. [16] Federal Aviation Administration, “System wide information management (SWIM) — overview,” https://www.faa.gov/air traffic/technology/ swim, 2025. [17] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: Synthetic minority over-sampling technique,” J. Artif. Intell. Res., vol. 16, pp. 321–357, 2002. [18] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2016, pp. 785–794. [19] L. Lamport, “The part-time parliament,” ACM Trans. Comput. Syst., vol. 16, no. 2, pp. 133–169, 1998. [20] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine Learning. MIT Press, 2006. [21] D. R. Jones, M. Schonlau, and W. J. Welch, “Efficient global optimization of expensive black-box functions,” J. Global Optim., vol. 13, no. 4, pp. 455–492, 1998. [22] M. D. McKay, R. J. Beckman, and W. J. Conover, “A comparison of three methods for selecting values of input variables in the analysis of output from a computer code,” Technometrics, vol. 21, no. 2, pp. 239– 245, 1979. [23] A. Dhodapkar, A. Smidt, A. Verkleeren, S. Patterson, and C. Varela, “Decentralized air traffic control (DATC) — source code,” https://github. com/Aditya-Dhodapkar/Decentralized-Air-Traffic-Control---DATC, 2025.
VI. F UTURE W ORK AND C ONCLUSION Future work. Submillisecond inference makes periodic resectorization feasible every few minutes as traffic patterns shift, with hysteresis to prevent oscillation. Testing under realistic communication models with packet loss and ADS-B latency distributions would quantify the gap between ideal and deployed performance, particularly the effect of packet loss on retry rates. Learning the BO objective weights from data via bilevel optimization or inverse reinforcement learning would remove the need for manual coefficient selection and could reveal that the relative importance of success rate, holding, speed modifications, and retries differs from what intuition suggests. Conclusion. We presented a three stage pipeline for adaptive airspace sectorization and decentralized admission control. A two stage XGBoost predictor selects optimal grid configurations at 91.38% accuracy from 23 location agnostic features. A Paxos consensus protocol coordinates sector entries with above 96% success and low NMAC rates across both sweeps. Bayesian Optimization automatically tunes protocol parameters, confirming that each airport requires its own configuration. The submillisecond prediction time and location-
6