RSS Evaluation Framework

• RSS Evaluation Framework is a comprehensive methodology that benchmarks localization accuracy by integrating parametric path-loss models with particle filtering techniques.
• It delineates experimental protocols that assess performance with metrics such as RMSE, convergence speed, and error variance across both good and bad sensor geometries.
• Empirical analysis demonstrates that particle filtering significantly improves robustness over traditional trilateration, especially in high-noise and poor geometric configurations.

A Received Signal Strength (RSS) evaluation framework provides the methodological and quantitative infrastructure for experimentally assessing RSS-based localization algorithms and systems. Such a framework delineates the measurement models, algorithmic approaches, noise models, experimental protocols, performance metrics, and design guidelines necessary for rigorously benchmarking localization accuracy and robustness, particularly under the practical realities of sensor geometry and signal variability. In the canonical formulation, RSS evaluation is essential for validating algorithms that estimate the position of a stationary target from distributed signal measurements, especially in applications such as wireless sensor networks, surveillance, and search and rescue.

1. Measurement Model and Signal Assumptions

The framework is grounded by a parametric path-loss model describing the physical relationship between the unknown target position $x\in\mathbb{R}^2$ and observed RSS values at $M$ sensors with known locations $s_m$:

$$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$

Key parameters:

• Reference RSS $P_0 = -30\,$dB at distance $d_0 = 1\,$m,
• Path-loss exponent $\beta = 2.5$,
• Additive measurement noise $v_m$ is i.i.d. Gaussian with standard deviation $\sigma$ ($\sigma$ varied between $M$0dB and $M$1dB to model different multipath/interference conditions).

The framework evaluates performance over both low- and high-noise regimes and systematically varies $M$2 to assess algorithmic robustness.

2. Particle Filtering Algorithmic Core

The primary algorithm under evaluation is a particle filtering approach explicitly tailored for stationary target localization. The state-space model is degenerate with process model $M$3 and thus no process noise. The measurement update for particle $M$4 at position $M$5 uses

$M$6

where $M$7 regularizes the log to avoid singularities at zero distance. Particles are initialized uniformly in the search area $M$8.

Importance weights for each particle at time $M$9:

$s_m$0

Weights are normalized, and the expected target estimate is taken as the mean of the (possibly resampled) particle set.

Resampling is controlled by a ratio $s_m$1 (default $s_m$2): the top $s_m$3 weighted particles are retained, and the remaining $s_m$4 replaced by new random samples.

Typical operational parameters are $s_m$5 particles, $s_m$6, $s_m$7, matching the above path-loss and noise settings.

3. Experimental Protocols and Sensor Geometries

Evaluation is performed over two fundamental sensor placement scenarios:

• Good geometry: Sensors arranged at the corners around the target, maximizing angular spread.
• Bad geometry: All sensors are co-located or clustered on one side of the target, embodying poor geometric diversity.

In all cases $s_m$8 sensors are used, and the search area is a $s_m$9 m square.

Each simulation run comprises $$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$0 independent epochs, with each epoch bringing a new independent RSS measurement. Error metrics are averaged over these to capture steady-state as well as transient convergence behavior, and Monte Carlo trials are repeated to estimate error variance.

4. Evaluation Metrics and Quantitative Criteria

Three main classes of metrics provide a comprehensive performance profile:

• Root-mean-square error (RMSE):

$$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$1

where $$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$2 is the true target position.

• Convergence speed: Number of epochs until RMSE falls below a fixed threshold (e.g., $$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$3 m).
• Error variance: Standard deviation of localization errors across stochastic trial repetitions.

These metrics support comparative analysis against traditional RSS-trilateration, whose solution is the least-squares fit to observed distances.

5. Key Findings and Analysis

Empirical results demonstrate critical dependencies on sensor geometry and noise:

• Good geometry: Both the particle filter (PF) and trilateration converge to RMSE $$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$4–$$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$5 m within $$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$6–$$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$7 epochs, with negligible performance difference.
• Bad geometry: The PF exhibits superior robustness: at $$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$8 dB after $$\hat{P}_m = P_0 - 10\beta \log_{10} \left( \frac{\| x - s_m \|}{d_0} \right) + v_m,\qquad v_m \sim \mathcal{N}(0,\sigma^2)$$9 epochs, PF achieves RMSE $P_0 = -30\,$0 m vs. $P_0 = -30\,$1 m for trilateration. Error variance is nearly halved for PF ($P_0 = -30\,$2 m vs. $P_0 = -30\,$3 m).
• Noise sweep: At $P_0 = -30\,$4 dB, methods are similar. For $P_0 = -30\,$5 dB, the PF’s non-Gaussian, nonlinear inference yields consistent performance advantages. At $P_0 = -30\,$6 dB, PF yields $P_0 = -30\,$7 m error, trilateration $P_0 = -30\,$8 m.
• Variance: PF exhibits $P_0 = -30\,$9–$d_0 = 1\,$0\% lower error variance under adverse conditions.

It is observed that PF’s sampling-based, nonlinear character provides strong resilience in ill-posed (clustered) sensor geometries and at high noise levels, regimes where analytic least-squares methods become highly unstable.

6. Design and Deployment Guidelines

The framework prescribes operational and deployment recommendations based on these results:

• Sensor placement: Favor maximum angular diversity; avoid clustering. PF is essential when clustering is unavoidable.
• Noise handling: PF is tolerable for $d_0 = 1\,$1 dB and RMSE $d_0 = 1\,$2 m; for higher noise, increase $d_0 = 1\,$3 and $d_0 = 1\,$4.
• Particle filter tuning: Use $d_0 = 1\,$5 for moderate fidelity, $d_0 = 1\,$6–$d_0 = 1\,$7 for $d_0 = 1\,$8 dB. Set $d_0 = 1\,$9, monitor effective sample size (ESS), and resample when $\beta = 2.5$0.

The framework provides a clear, replicable recipe for practitioners evaluating RSS-based localization in realistic or adversarial deployment scenarios. Its prescriptions are robust over a wide noise range and geometric spectrum and can be leveraged to guide both experimental studies and fielded system design, ensuring that solutions are optimized not merely for ideal cases but for the practical complexities of wireless sensor network operation (Lee et al., 19 Aug 2025).