Gai Nerf Eval

---
name: gai-nerf-eval
description: Evaluates the accuracy and generalization of wireless channel prediction models across diverse indoor environments and frequency bands. It probes the model's ability to predict received signal strength (RSSI) and channel state information (CSI) given spatial coordinates and environmental geometry, while testing robustness to physical scene changes and cross-frequency translation. Use when the user wants to benchmark on Our Own Datasets, Argos channel dataset, NewRF simulated data, or asks about evaluating this task. Reports MAE (dB), SNR (dB).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.11983
  bibtex_key: shen2026gai-nerf
  confidence: high
---

# gai-nerf-eval

> A Geometric Algebra-informed NeRF Framework for Generalizable Wireless Channel Prediction — Shen et al. (2026) (arXiv:2604.11983, 2026)

## What this evaluates

Evaluates the accuracy and generalization of wireless channel prediction models across diverse indoor environments and frequency bands. It probes the model's ability to predict received signal strength (RSSI) and channel state information (CSI) given spatial coordinates and environmental geometry, while testing robustness to physical scene changes and cross-frequency translation.

## Datasets

- **Our Own Datasets** — total ?; splits: Room 1 (-1), Room 2 (-1)
- **Argos channel dataset** — total 100000; splits: (unstated)
- **NewRF simulated data** — total 443975; splits: Bedroom (443975), Conference (443975)

## Metrics

- `MAE (dB)` **(primary)** — range: dB
  - Mean Absolute Error calculated as the average of absolute differences between predicted and ground-truth signal values. Results are aggregated as medians across frequencies and rooms.
- `SNR (dB)` **(primary)** — range: dB
  - Signal-to-Noise Ratio calculated as 10 * log10(P_signal / P_noise), where P_signal is the mean squared predicted value and P_noise is the mean squared error. Results are aggregated as medians.

## Input / output format

**Input**: Spatial coordinates (x, y, z), environmental geometry (e.g., occupancy grids, ray paths), and operating frequency band (2.4 GHz or 5 GHz).

**Output**: Predicted Received Signal Strength Indicator (RSSI) in dB or Channel State Information (CSI) values.

## Scoring recipe

```python
def compute_metrics(predictions, ground_truth):
    mae = np.mean(np.abs(predictions - ground_truth))
    signal_power = np.mean(predictions**2)
    noise_power = np.mean((predictions - ground_truth)**2)
    snr_db = 10 * np.log10(signal_power / noise_power)
    return mae, snr_db
# Aggregate results as medians across frequencies/rooms as specified in the paper.
```

## Common pitfalls

- GWRF and NewRF baselines are excluded from RSSI evaluations because their pooling/sampling mechanisms compress variance, causing them to fail on low-variance RSSI data.
- MAE and SNR are reported in decibels (dB), not linear power units.
- Generalization is tested via cross-frequency translation (2.4 GHz ↔ 5 GHz) and physical scene modifications (adding/relocating/removing objects), not just standard train/test splits.

## Evidence (verbatim from paper)

> We evaluate performance using the median of MAE and signal-to-noise ratio (SNR) as our primary metrics.

## Citation

```bibtex
@misc{shen2026gai-nerf,
  title={A Geometric Algebra-informed NeRF Framework for Generalizable Wireless Channel Prediction},
  author={Shen et al. (2026)},
  year={2026},
  note={arXiv:2604.11983}
}
```

- arXiv: 2604.11983