Improving D-Optimal Sensor Placement for Bearing-Only Localization via Maximum-Entropy Reweighting

Technical Summary: Improving D-Optimal Sensor Placement for Bearing-Only Localization via Maximum-Entropy Reweighting

Problem Statement
The paper addresses the challenge of optimal sensor placement for bearing-only source localization, a problem prevalent in sonar, radar, and wildlife tracking. The standard approach utilizes D-optimal design, which maximizes the log-determinant of the Fisher Information Matrix (FIM). However, in particle-filter implementations, the FIM is typically evaluated using uniformly weighted particles after resampling. This uniform averaging treats particles near the true source and those far away with equal importance, particularly in early iterations when the prior distribution spans a broad domain. Consequently, the resulting sensor placement is optimal in an expected sense over the prior but fails to exploit the specific geometric configurations where the posterior mass is likely to concentrate. This limitation is exacerbated in multi-source scenarios where the joint state space is high-dimensional and the posterior structure is complex.

Methodology
The authors propose a two-layer sequential architecture that decouples the definition of information requirements from the sensor configuration optimization:

1. Layer 1: Maximum-Entropy (MaxEnt) Reweighting:
   Before optimizing sensor positions, the particle weights are reweighted to concentrate mass on regions where the posterior is likely to settle, without referencing the specific sensor model. This is formulated as a constrained optimization problem: minimize the Kullback-Leibler (KL) divergence between the new weights and the current prior weights, subject to a distributional accuracy bound defined by the Sliced Wasserstein (SW2) distance.

   • Target: The target distribution is a Dirac mass at the current weighted mean of the particles.
   • Constraint: The SW2 distance between the reweighted empirical measure and the target must be within a budget $\epsilon$.
   • Solution: Because the target is a Dirac mass, the SW2 constraint becomes linear in the weights. The solution is an exponential tilt of the prior weights: $w_i^{opt} \propto w_i^- \exp(-\lambda g_i)$, where $g_i$ represents the cost (squared distance) of particle $i$ relative to the target. The Lagrange multiplier $\lambda$ is determined via bisection to satisfy the accuracy budget.
   • Decoupling: This layer is independent of sensor noise, count, or position, making the reweighting generalizable across different sensing modalities.

2. Layer 2: D-Optimal Sensor Placement:
   Using the reweighted particles from Layer 1, the system solves the standard D-optimal placement problem. The objective is to maximize the log-determinant of the total FIM, which is now computed using the reweighted particles. This steers sensors toward geometrically informative configurations specific to the concentrated posterior regions. The optimization employs multi-start L-BFGS-B with analytically computed gradients.

3. Sequential Procedure:
   The process iterates: Layer 1 reweights particles, Layer 2 optimizes sensor positions, synthetic bearing measurements are generated, and particle weights are updated via standard Bayesian multiplication. Resampling occurs when the Effective Sample Size (ESS) drops below a threshold.

Key Contributions

1. Framework Instantiation: The paper instantiates the maximum-entropy likelihood synthesis framework for multi-source bearing-only localization, combining MaxEnt particle reweighting with D-optimal placement in a sequential architecture.
2. Systematic Empirical Evidence: The authors provide experimental results across 64 configurations (varying source counts $M \in \{3, \dots, 10\}$ and sensor counts $R \in \{3, \dots, 10\}$) at two noise levels ($\sigma = 8^\circ$ and $\sigma = 15^\circ$).
3. Regime Characterization: The paper identifies the specific conditions under which the method succeeds or fails. It characterizes the "success regime" as having a sensor-to-source ratio $R > M$ and informative measurements (low noise). Conversely, it identifies a "failure regime" where $R < M$ (sensor-starved), noting that reweighting can degrade performance by sacrificing necessary angular diversity.

Results

• Performance Improvement: At moderate noise ($\sigma = 15^\circ$), the MaxEnt method reduced the Root-Mean-Square Error (RMSE) by a mean of 7.9% compared to the D-optimal baseline, with improvements ranging from -21% to +25%. At low noise ($\sigma = 8^\circ$), the mean improvement rose to 17.2%, with a range of -7% to +34%.
• Sensor Efficiency: The method demonstrated significant "sensor savings." At $\sigma = 8^\circ$, MaxEnt with fewer sensors often achieved the same RMSE as D-optimal with many more sensors. For instance, with 6 sources, MaxEnt using 3 sensors outperformed D-optimal using 10 sensors, representing a saving of 7 sensors.
• Convergence: The improvement is established in the first few iterations of the sequential procedure. As the posterior concentrates, the reweighted and uniform distributions converge, but the early advantage in sensor placement persists, leading to a lower final RMSE.
• Failure Modes: The method degraded performance primarily when sensors were scarce relative to sources ($R < M$). In these cases, the reweighting concentrated the FIM on a subset of the posterior, reducing the angular diversity required for accurate localization.

Significance and Claims
The paper claims that the proposed two-layer architecture offers a principled way to improve sensor placement by aligning the FIM evaluation with the likely geometry of the posterior, rather than the broad prior. The structural decoupling is highlighted as a key theoretical advantage, allowing the reweighting logic to generalize across sensing modalities (e.g., range, time-of-arrival) while the placement layer remains specific to the geometry.

The authors modestly frame their contribution as an improvement over classical D-optimal design under specific conditions, rather than a universal solution. They explicitly acknowledge limitations, including the assumption of known data association (which avoids the combinatorial complexity of joint probabilistic data association), the restriction to 2D static sources, and the use of a Dirac-at-mean target which simplifies the optimization but may not capture multi-modal posteriors as effectively as Gaussian mixture targets. The work establishes that while the method is provably beneficial when geometric freedom exists ($R > M$) and measurements are informative, it reverts to classical D-optimal behavior or degrades performance when these conditions are not met.