Search strategy in a complex and dynamic environment: the MH370 case

The paper addresses the problem of planning efficient aerial and maritime surface search operations, exemplified by the MH370 incident, where extensive efforts failed to locate debris. Two key challenges hinder ocean-surface search: (1) high uncertainty in target area estimation due to evolving and conflicting information, compounded by complex, unsteady ocean drift that amplifies splash-location estimation errors; and (2) the lack of known optimal path solutions for search agents even when a probabilistic target distribution is available. Conventional lawnmower sweeps are easy to implement in simple geometries but are ill-suited to the complex, non-uniform, and dynamically evolving shapes produced by ocean drift, and they lack operational flexibility. The study proposes a multi-agent motion control algorithm, modified Dynamic Spectral Multi-scale Coverage (mDSMC), built on ergodic theory and optimal search, designed to handle complex geometries, non-uniform likelihoods, and target drift, with the goal of improving detection probability in dynamic ocean environments. The MH370 case is used to evaluate the approach against conventional methods.

Related research is grouped into stationary and non-stationary target search. For stationary targets: UAV route planning over rivers uses Gaussian mixture models with heuristic prioritization and receding horizon control (RHC/MPC) for single and multi-UAV scenarios; ergodic exploration has been posed via optimization; energy-constraint-aware planners, advanced lawnmower variants considering vehicle, sensor, and partial-detection models, and hierarchical GMM-driven coverage improve sweeps. Swarm path planning leverages particle swarm optimization with mobility and communication constraints; reinforcement learning has been used for indoor SAR via RF signals; heat-equation-driven area coverage (HEDAC) and exact probabilistic models achieve near-optimal stationary target performance; and multi-agent layered SAR strategies aim to minimize time while maximizing finds. For non-stationary targets: multi-UAV minimum-time search employs Ant Colony Optimization and variants using target probability maps; hybrid evolutionary and Tabu search approaches evolve multi-UAV plans under time-varying target probabilities; distributed cooperative search uses predictions and RHC to maximize performance; ergodic exploration with RHC plans motion against dynamic information; game-theoretic approaches consider evasive targets. Mission planners incorporate weather and icing for ocean debris search; market-based and evolutionary strategies reduce cost. Coordinated UAV-AUV search has been studied; AUV search in unsteady currents uses self-organizing maps and velocity synthesis. Search-theoretic methods coupled with surrogate ocean models address drifting object localization. Multi-UAV motion/camera control minimizes detection time for nondeterministically moving targets from uncertain initial positions; reinforcement learning has been applied for moving ships at sea. Foundational works include ergodicity metrics and spectral multi-scale coverage (SMC), its extension to dynamic environments (DSMC), and classic optimal search theory (Koopman, Stone). Prior MH370-related drift and dynamical systems analyses inform drift modeling and convergence zone concepts.

Core framework: The search domain is a large rectangular ocean-surface area. Two key fields define the problem: (1) a time-evolving target probability distribution p(x,t) transported by the surface velocity field v(x,t) (assumed incompressible), and (2) a coverage distribution capturing where and how long agents have searched. The target distribution evolves via advection; in simulations a Lagrangian, semi-Lagrangian approach is used to avoid artificial diffusion. Coverage incorporates sensor characteristics via a smooth radial basis function (RBF) to model detection density around agents. Detection follows an exponential saturation model F(c) = 1 − exp(−c), consistent with optimal search theory.

Optimal coverage: For stationary p, Koopman’s theory yields an optimal coverage distribution c_opt = max[ln p − α, 0], where α is chosen to satisfy the total available coverage constraint over the time window. In dynamic settings and incremental planning, the same principle applies with updated posterior p.

Mismatch/objective for control: Instead of directly optimizing the nonconvex detection probability over paths, the algorithm minimizes the mismatch between current coverage and the Koopman-optimal distribution in a Sobolev negative-index norm that emphasizes large-scale features. The modified DSMC (mDSMC) defines mismatch s^t = max[ln p^t − α, 0], explicitly excluding over-searched areas (ln p^t < α), improving local search and stability over DSMC’s simpler mismatch s^t = p^t − c^t/N'. The mismatch norm employs Fourier coefficients weighted by λ_k = (1 + ||k||^2)^β with β negative; mDSMC uses β = −1/2 (less smoothing, better late-stage fine-scale exploration) while DSMC uses β = −3/2.

Control law: A potential field is constructed from the weighted mismatch spectrum. Each agent moves with first-order dynamics at constant speed along the steepest descent direction of the potential, normalized. The RBF used is a compact-support approximation of a normal distribution centered at the agent with standard deviation σ = 3 km. A 1-hour window is used to determine available coverage for Koopman allocation.

MH370 simulation scenarios: Four scenarios are considered: (1) Lawnmower scenario 1 replicates the reported JACC/ATSB search areas and schedules (areas A and B, daily aircraft counts as reported), (2) Lawnmower scenario 2 drifts ATSB splash areas with the study’s drift model and searches the convex hull of the resulting support by lawnmower, (3) DSMC searches the drifted distribution using DSMC, and (4) mDSMC searches the drifted distribution using mDSMC. Scenario 1 represents actual operations; scenarios 2–4 isolate algorithmic performance under a common drift model.

Drift and detection models: Surface velocity data are from HYCOM (US Navy) with 3-hourly fields at 1/25° resolution. Targets are initialized at 00:30 UTC on March 8 using a Halton sequence and advected with a 4th-order adaptive Runge–Kutta method with linear interpolation in space and time. Each splash area is sampled with 10^5 tracers to build p(x,t) via local density of advected tracers. Detection is modeled stochastically: if a target remains within 1.5 km of an agent for time t, the detection probability is P = 1 − exp(−t/T) with expected detection time T = 2 s, approximating human visual detection as used during MH370.

Agent and search parameters: Only aircraft are modeled as agents due to unavailable ship specifications. Agent speed is 380 km/h (typical loiter speed). Each search day simulates 3 hours of on-station scanning (approximate transit accounted for), from 14:00–17:00 UTC. Monte Carlo evaluation seeds 1000 targets randomly in the splash area on the splash day and drifts them; each scenario is run over 100 realizations with random initial target and agent positions to compute ensemble success rates. Areas considered: A (searched Mar 28–Apr 1), B (Apr 2–3), and P (identified later as most probable but not surface-searched; used for delayed-start and uncertainty analyses).

Delayed start analysis: For area P, mDSMC searches commencing 0, 5, and 10 days after splash are compared to study how surface mixing and convergence zones affect early-stage detection.

Uncertainty in drift: To assess robustness to flow-field errors, target drift is augmented with Brownian motion via a Langevin SDE dy = v(y) dt + B dW, where B relates to an effective diffusion (isotropic standard deviations V_err ∈ {0.1, 0.2, 0.5, 1} m/s). In practice, random displacements are applied during advection to mimic uncertainty; performance across 100 runs is compared per uncertainty level.

• Algorithmic performance on area A (end of search period, single-run visualizations):
  • Lawnmower scenario 1 (replicated reported areas): detection rate 33.0%.
  • Lawnmower scenario 2 (drifted support, convex-hull sweep): 46.3%.
  • DSMC: 65.9%.
  • mDSMC: 93.5%.
  • Across 100-run ensembles (histogram means, where available): lawnmower scenario 1 mean ≈ 33.84%; lawnmower scenario 2 mean ≈ 43.77%; mDSMC mean ≈ 92.45%.
• Algorithmic performance on area B (end of search period, single-run visualizations):
  • Lawnmower scenario 1: 6.6%.
  • Lawnmower scenario 2: 11.7%.
  • DSMC: 28.2%.
  • mDSMC: 55.7%.
  • Across 100-run ensembles (means): lawnmower scenario 1 ≈ 10.89%; lawnmower scenario 2 ≈ 10.36%; DSMC ≈ 24.87%; mDSMC ≈ 55.78%.
• Qualitative trajectory behavior: mDSMC adapts to thin, filamentary ribbons of drifted probability, avoids over-searched regions by construction, and balances early broad sweeps with later localized search; DSMC tends toward broader, less locally refined coverage; lawnmower sweeps are inefficient for complex, indented shapes and may over-cover zero-probability regions or miss significant probability mass.
• Delayed start on area P: mDSMC simulations indicate higher initial-day detection rates when the search commences 5–10 days after splash, due to convergence zones that accumulate debris and reduce effective search area. Daily detection-rate differences are positive in early days for delayed starts, diminishing over time. The study emphasizes this observation does not advocate delaying SAR for survivors.
• Robustness to drift uncertainty (area P, 100-run ensembles; means, standard deviations):
  • Deterministic drift (V_err = 0 m/s): mean ≈ 84.55% (σ ≈ 1.22).
  • V_err = 0.1 m/s: mean ≈ 81.49% (σ ≈ 1.51).
  • V_err = 0.2 m/s: mean ≈ 78.08% (σ ≈ 1.11).
  • V_err = 0.5 m/s: mean ≈ 71.46% (σ ≈ 1.37). Despite deteriorating performance with higher uncertainty, mDSMC maintains substantially higher success than lawnmower strategies.
• Resource/schedule realism: Simulations limit daily on-station time to 3 hours and use reported aircraft counts and periods, reinforcing that gains arise from path planning rather than increased effort.

The findings address the central research question of how to plan search paths in dynamic, uncertain ocean environments to maximize detection probability. By leveraging Koopman’s optimal search theory to define a target-aligned optimal coverage and minimizing a Sobolev-weighted mismatch that de-emphasizes small scales while excluding over-searched regions, mDSMC guides agents along the large-scale structures of the drifted probability distribution and subsequently refines searches in finer-scale filaments. This results in significantly higher detection rates than conventional lawnmower sweeps and consistent improvements over DSMC, which lacks the explicit exclusion of over-searched areas and employs stronger smoothing. The MH370 replication shows order-of-magnitude improvement in success rate over the same time horizon. The delayed-start analysis provides insight into dynamic-environment search: convergence zones revealed by finite-time Lagrangian hypergraph analysis concentrate targets, initially increasing encounter rates if search starts later. This mechanism explains higher initial detection when starting 5–10 days after splash in the simulations. However, the practical imperative in SAR to find survivors quickly precludes intentional delay; rather, the insight suggests deploying algorithms aware of convergence structures to prioritize regions of early accumulation. Under drift uncertainty, performance degrades gracefully as target dispersion increases and mismatches the predicted probability density. Nonetheless, mDSMC’s emphasis on large-scale coherent structures preserves much of its advantage. Overall, the results demonstrate that incorporating flow-aware, ergodic-inspired coverage control can substantially enhance operational search effectiveness in complex, time-varying environments.

The study introduces mDSMC, a flow-aware, ergodic-theory-based multi-agent search algorithm that plans paths by minimizing a Sobolev-norm mismatch to the Koopman-optimal coverage. Compared with conventional lawnmower sweeps and the prior DSMC method, mDSMC achieves markedly higher detection rates in realistic simulations of the MH370 surface search, handling complex, filamentary probability distributions and dynamically evolving drift. Monte Carlo experiments, delayed-start analyses, and robustness tests under drift uncertainty all confirm its effectiveness and operational promise. Potential future research directions include: integrating higher-fidelity and data-assimilative ocean models and real-time drifter/satellite data; jointly planning aerial and maritime (ship and AUV) assets with heterogeneous sensing; extending to second-order agent dynamics and energy/resource constraints; incorporating adaptive sensing and online probability updating with detections and false alarms; addressing multi-objective trade-offs such as survivor rescue vs. debris recovery; and efficient real-time implementations for large fleets using distributed computation and communication constraints.

• Dependence on accurate ocean surface velocity fields: performance relies on the fidelity of HYCOM or measured flows; uncertainties degrade results as shown.
• Simplified detection model: human visual detection is approximated with a uniform 1.5 km radius and an exponential saturation with T = 2 s; real detection depends on visibility, sea state, and observer workload.
• Ships excluded: only aircraft were modeled due to lack of ship specifications, omitting potential contributions and coordination with maritime assets.
• Advection-only probability evolution in baseline: full advection-diffusion PDE was not solved due to computational cost; uncertainty was modeled via stochastic perturbations rather than full PDE-based diffusion of p.
• Use of convex hull for lawnmower scenario 2 may advantage/disadvantage comparisons depending on true support geometry; real operations may use more complex area assignments.
• The delayed-start benefit is context-specific to debris accumulation and does not apply to survivor rescue where immediate search is critical; sinking or degradation of targets over time was not modeled.
• Agent dynamics simplified to first-order constant speed; constraints such as fuel, maneuverability, and communication limits were not fully modeled.