Milan Bs Sleep Eval

---
name: milan-bs-sleep-eval
description: Evaluates a deep reinforcement learning framework for dynamic base station sleep control and spatio-temporal traffic forecasting in a real-world cellular network. It probes the model's ability to accurately predict mobile traffic demand across geographical grids and make energy-efficient on/off decisions for base stations while balancing switching costs and quality of service. Use when the user wants to benchmark on Telecom Italia Milan Mobile Traffic Dataset, or asks about evaluating this task. Reports NMAE.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2101.08391
  bibtex_key: wu2021deepbsc
  confidence: high
---

# milan-bs-sleep-eval

> Deep Reinforcement Learning with Spatio-temporal Traffic Forecasting for Data-Driven Base Station Sleep Control — Wu et al. (2021) (arXiv:2101.08391, 2021)

## What this evaluates

Evaluates a deep reinforcement learning framework for dynamic base station sleep control and spatio-temporal traffic forecasting in a real-world cellular network. It probes the model's ability to accurately predict mobile traffic demand across geographical grids and make energy-efficient on/off decisions for base stations while balancing switching costs and quality of service.

## Datasets

- **Telecom Italia Milan Mobile Traffic Dataset** — total ?; splits: train (-1), test (-1)

## Metrics

- `NMAE` **(primary)** — range: [0, 1]
  - Normalised Mean Absolute Error: NMAE = (1/d_bar) * sum(|d_hat_k - d_k| / N), where d_hat_k is predicted traffic, d_k is ground truth, N is number of grids, and d_bar is mean traffic.
- `NRMSE` — range: [0, 1]
  - Normalised Root Mean Square Error: NRMSE = (1/d_bar) * sqrt(sum((d_hat_k - d_k)^2 / N)).
- `Normalized system cost` — range: percent
  - Aggregated cost metric combining energy consumption, base station switching costs, and QoS degradation, normalized against a baseline traffic-oblivious policy.

## Input / output format

**Input**: For forecasting: historical mobile traffic measurements over the previous 48 time slots (30-min intervals) across 100 geographical grids. For sleep control: a 200-dimensional state vector comprising traffic demand in 100 grids and current operation modes of 100 base stations.

**Output**: For forecasting: predicted mobile traffic demand per grid for the next time step. For sleep control: binary action vector (100 dimensions) indicating active or sleep mode for each base station.

## Scoring recipe

```python
import numpy as np
def compute_metrics(d_hat, d):
    N = len(d)
    d_bar = np.mean(d)
    nmae = (1 / d_bar) * np.sum(np.abs(d_hat - d) / N)
    nrmse = (1 / d_bar) * np.sqrt(np.sum((d_hat - d)**2 / N))
    return nmae, nrmse
# Normalized system cost = (Total Cost of Method) / (Total Cost of Traffic-Oblivious Baseline)
```

## Common pitfalls

- Aggregating 10-minute interval data to 30-minute intervals without proper alignment can distort traffic patterns and mislead the model.
- Ignoring semantic correlations between non-adjacent grids with similar traffic patterns leads to suboptimal forecasting compared to purely spatial models like CNN-LSTM.
- High variance in cost estimation due to traffic fluctuations can mislead DRL training unless benchmark transformation is applied.

## Evidence (verbatim from paper)

> To quantify the performance of the proposed GS-STN method and existing prediction methods, we use the Normalised Mean Absolute Error (NMAE) and Normalised Root Mean Square Error (NRMSE) as given below: NMAE = (1/d_bar) sum(|d_hat_k - d_k|/N), NRMSE = (1/d_bar) sqrt(sum((d_hat_k - d_k)^2/N))... We train our DeepBSC framework on the dataset during the first 20 days and evaluate on the last 10 days.

## Citation

```bibtex
@misc{wu2021deepbsc,
  title={Deep Reinforcement Learning with Spatio-temporal Traffic Forecasting for Data-Driven Base Station Sleep Control},
  author={Wu et al. (2021)},
  year={2021},
  note={arXiv:2101.08391}
}
```

- arXiv: 2101.08391