SALAD: Self-Adaptive Link Adaptation

1. Problem Statement

Link Adaptation (LA) is critical for maximizing spectral efficiency (SE) in wireless networks by dynamically selecting the Modulation and Coding Scheme (MCS) based on channel conditions. The core challenge is balancing throughput (higher MCS) with reliability (low Block Error Rate, BLER).

• Current Standard (OLLA): The industry standard, Outer-Loop Link Adaptation (OLLA), uses a fixed stepsize to adjust an SINR offset based on ACK/NACK feedback.
  • Limitation: A fixed stepsize creates a trade-off. A small stepsize adapts too slowly to rapid channel changes (hurting short-lived connections), while a large stepsize causes instability and oscillation in static conditions, degrading SE.
• SINR Estimation Issues: Accurate SINR estimation is difficult. Channel State Information (CQI) reports are often coarse (integer 0–15), infrequent, delayed, and wide-band, making them unreliable for precise, real-time MCS selection.
• Goal: Develop an algorithm that can rapidly track SINR variations without sacrificing long-term BLER stability, reducing reliance on imperfect CQI reports.

2. Methodology: SALAD Algorithm

The authors propose SALAD (Self-Adaptive Link Adaptation), an algorithm that infers SINR exclusively from ACK/NACK feedback and adapts its behavior dynamically. It consists of three core modules:

A. SINR Inference via Cross-Entropy Minimization

Unlike OLLA, which uses a fixed additive offset, SALAD treats SINR estimation as an online optimization problem.

• Mechanism: It minimizes the Binary Cross-Entropy (BCE) loss between observed ACK/NACK outcomes and predicted BLER values derived from a sigmoid approximation of the BLER curve.
• Update Rule: The SINR estimate ($\hat{\gamma}_t$) is updated via gradient descent:
  $$\hat{\gamma}_t = \hat{\gamma}_{t-1} + \varepsilon \cdot s^{-1} \cdot (\text{BLER}_{t-1} - \mathbb{1}_{\text{NACK}})$$
  Where $\varepsilon$ is the learning rate and $s$ is a scale parameter.
• "Surprise Effect": The update magnitude depends on the discrepancy between expectation and reality.
  • If an ACK is received when a high BLER was expected (SINR underestimated), the SINR estimate jumps up significantly.
  • If a NACK is received when a low BLER was expected (SINR overestimated), the estimate drops significantly.
  • This allows SALAD to correct estimation errors much faster than OLLA's fixed steps.

B. MCS Selection via Hypothesis Testing (Probing)

To accelerate adaptation when the channel improves, SALAD actively probes higher MCS values.

• Bias Score: It calculates a "bias score" ($S_t$) over a sliding window, representing the cumulative difference between predicted BLER and actual ACK/NACKs.
• Hypothesis Test: If the normalized bias score exceeds a threshold ($\rho$), the algorithm rejects the null hypothesis that the SINR estimate is correct. It infers that the SINR is underestimated.
• Probing: Upon detecting underestimation, SALAD temporarily raises the instantaneous BLER target ($\tau_{probe} \gg \tau$) to select a more aggressive MCS.
  • If successful (ACK), the "surprise effect" rapidly corrects the SINR estimate upward.
  • If it fails (NACK), the integral control loop (below) compensates.

C. Long-Term BLER Enforcement (Feedback Control)

To prevent the aggressive probing from causing the long-term BLER to drift above the target:

• Integral Control: SALAD maintains an integral error term ($E_t$) of the difference between the target BLER and actual outcomes.
• Dynamic Target Adjustment: The instantaneous BLER target ($\tau_t$) is adjusted based on this integral error (similar to a PID controller). If the BLER exceeds the target, $\tau_t$ is lowered to force a more conservative MCS selection, ensuring the long-term average remains at the desired level (e.g., 10%).

3. Key Contributions

1. Principled SINR Inference: A novel method to infer SINR using only ACK/NACK feedback by minimizing cross-entropy loss, eliminating the need for coarse/delayed CQI reports.
2. Adaptive Stepsize via Probing: Instead of a fixed stepsize, SALAD uses hypothesis testing to detect SINR underestimation and dynamically "probes" higher MCS values, creating a virtuous cycle for rapid adaptation.
3. Robustness to Delay: The algorithm is mathematically shown to be more robust to HARQ feedback delays compared to OLLA because it updates based on effective BLER rather than a fixed offset rule.
4. Single-Set Parameterization: SALAD achieves high performance across diverse traffic regimes (low to high throughput) with a single set of parameters, whereas OLLA requires different stepsize tuning for different scenarios.

4. Experimental Results

The authors validated SALAD using full-stack system simulations (OAI RFSim) and real-world Over-the-Air (OTA) experiments in an anechoic chamber using a 5G testbed (NVIDIA Aerial + OAI).

• Throughput Gains: SALAD achieved up to 15% higher throughput and spectral efficiency compared to multiple OLLA variants (with different fixed stepsizes) across various traffic regimes (UDP 1/10 Mbit/s and TCP 100 Mbit/s).
• Adaptation Speed: In scenarios with sudden channel improvements (SINR surges), SALAD adapted approximately twice as fast as OLLA.
• Robustness to Delay: In experiments with delayed CQI reports, SALAD recovered from SINR undershoots significantly faster than OLLA.
• BLER Stability: Despite aggressive probing, SALAD maintained the long-term BLER close to the 10% target. Short-term BLER fluctuations were slightly higher during probing but were quickly corrected by the integral control loop.
• Parameter Tuning: While SALAD performs well with default parameters, the authors showed that derivative-free optimization (Nelder-Mead) could yield an additional ~3% throughput gain in specific regimes, though the default "self-adaptive" nature was already highly effective.

5. Significance

• Beyond OLLA: SALAD represents a shift from heuristic, fixed-step adaptation to a data-driven, optimization-based approach that learns the channel state directly from transmission outcomes.
• 5G/6G Relevance: As networks move toward higher frequencies and more dynamic environments (e.g., high mobility, URLLC), the limitations of fixed-step OLLA become more pronounced. SALAD offers a robust solution that does not rely on the often-imperfect CQI feedback mechanisms.
• Practical Implementation: The algorithm is lightweight, requires no complex side-information, and has been implemented in a real-time 5G stack, demonstrating immediate applicability to Open RAN (O-RAN) and commercial deployments.

In summary, SALAD successfully combines the theoretical rigor of stochastic approximation with practical adaptive mechanisms (probing and integral control) to outperform the industry standard in both speed of adaptation and overall spectral efficiency.