Link Adaptation

The closed-loop mechanism where user equipment (UE) measures downlink channel conditions and reports a CQI index to the base station. The scheduler maps this 4-bit value to a specific MCS entry.

• CQI reporting can be periodic (every 2-160 ms) or aperiodic (triggered by DCI)
• Wideband CQI reports a single value for the entire bandwidth; sub-band CQI provides per-resource-block granularity
• The UE estimates SINR and selects the highest CQI that achieves ≤10% BLER under current conditions
• 5G NR supports up to 31 CQI table entries for 256QAM and MIMO configurations

The base station maps reported CQI to an MCS index that determines the modulation order and code rate for the physical downlink shared channel (PDSCH).

• QPSK (2 bits/symbol): Robust for poor SINR, low data rate
• 16QAM (4 bits/symbol): Moderate throughput for medium channel conditions
• 64QAM (6 bits/symbol): High throughput for good SINR
• 256QAM (8 bits/symbol): Maximum spectral efficiency, requires excellent SINR >20 dB
• Code rates range from ~0.08 (high redundancy) to ~0.93 (minimal redundancy)
• Outer loop link adaptation adjusts MCS selection based on HARQ ACK/NACK statistics to correct systematic CQI measurement errors

Link adaptation operates in two nested control loops with different time constants and objectives.

• Inner loop: Fast adaptation (every TTI, 1 ms) based on instantaneous CQI reports. Selects MCS directly from the CQI-to-MCS mapping table
• Outer loop: Slower correction (every 10-100 ms) that monitors HARQ statistics to adjust the SINR-to-CQI mapping offset
• If BLER exceeds target, the outer loop applies a negative offset to make MCS selection more conservative
• If BLER is below target, a positive offset increases aggressiveness to improve throughput
• This dual-loop structure compensates for CQI estimation errors, Doppler effects, and UE receiver implementation variations

Pre-defined lookup tables translate estimated Signal-to-Interference-plus-Noise Ratio (SINR) to the optimal MCS that maximizes throughput while satisfying the target BLER constraint.

• Each MCS entry has a minimum SINR threshold for reliable decoding
• Tables are generated through link-level simulations using AWGN or fading channel models
• 5G NR defines multiple MCS tables optimized for different use cases: Table 1 (up to 64QAM), Table 2 (up to 256QAM), Table 3 (low-SE for URLLC)
• The mapping is non-linear: SINR improvements yield diminishing throughput gains at higher modulation orders
• Adaptive MCS tables can be tuned per cell based on long-term propagation environment characteristics

Traditional CQI-to-MCS mapping relies on fixed tables that cannot adapt to non-stationary channel dynamics. Deep reinforcement learning (DRL) agents learn adaptive MCS selection policies directly from interaction with the wireless environment.

• The state space includes CQI history, HARQ feedback, Doppler estimates, and buffer status
• The action space is the discrete set of MCS indices available for selection
• The reward function balances throughput maximization against BLER penalties: R = throughput - α × BLER_violation
• DRL agents can learn to anticipate fast fading dips and proactively reduce MCS before packet loss occurs
• Proximal Policy Optimization (PPO) and Deep Q-Networks (DQN) are commonly used algorithms
• DRL-based adaptation has demonstrated 15-25% throughput gains over conventional OLLA in high-mobility scenarios

Modern link adaptation integrates Hybrid Automatic Repeat Request (HARQ) feedback to dynamically adjust the first-transmission BLER target based on the reliability provided by retransmissions.

• If HARQ retransmissions are available with low latency, the initial transmission can target a higher BLER (e.g., 30%) to boost spectral efficiency
• The effective throughput accounts for the probability of retransmission: Effective Rate = Initial Rate × (1 - BLER) + Retransmission Rate × BLER × (1 - BLER_retx)
• Incremental redundancy HARQ provides additional parity bits on retransmission, improving decoding probability
• HARQ-aware adaptation is critical for URLLC services where latency budgets may allow only 1-2 retransmissions
• The outer loop offset can be conditioned on the HARQ process ID to track per-process reliability

The direct output of the link adaptation algorithm. The MCS index defines a specific combination of modulation order and code rate from a standardized table.

• Modulation Order: Ranges from QPSK (2 bits/symbol) for poor channel conditions to 256QAM (8 bits/symbol) for excellent SINR.
• Code Rate: The ratio of information bits to total transmitted bits. Lower code rates (e.g., 1/3) add more redundancy for error correction at the cost of throughput.
• MCS Tables: 5G NR defines multiple tables (e.g., Table 5.1.3.1-1 for up to 64QAM, Table 5.1.3.1-2 for 256QAM) to support different spectral efficiency targets.

Selecting an overly aggressive MCS results in decoding failures and retransmissions, while an overly conservative MCS wastes spectral resources.

A corrective control loop that compensates for systematic errors in CQI reporting. The UE's CQI estimate is inherently imperfect due to measurement noise, mobility, and reporting delay.

• Mechanism: OLLA maintains a dynamic offset (Δ_offset) applied to the reported SINR before MCS selection.
• Adjustment Logic: If a Cyclic Redundancy Check (CRC) passes, Δ_offset is increased by a small step-up value, making the link more aggressive. If a CRC fails, Δ_offset is decreased by a larger step-down value.
• Target BLER: The ratio of step-up to step-down sizes determines the converged BLER (e.g., a 10% target BLER requires step_down = 9 × step_up).

OLLA is essential for maintaining a stable BLER in the face of CQI estimation inaccuracies without requiring explicit channel prediction models.

The fundamental physical-layer metric that governs the achievable spectral efficiency. SINR is the ratio of the received power of the desired signal to the sum of interference power from other cells and thermal noise.

• SINR-to-MCS Mapping: Link adaptation relies on pre-computed lookup tables or curves that map SINR to the maximum MCS achieving the target BLER under Additive White Gaussian Noise (AWGN) conditions.
• Effective SINR: In OFDM systems with frequency-selective fading, an effective SINR metric (e.g., Exponential Effective SINR Mapping, EESM) compresses per-subcarrier SINRs into a single value for MCS selection.
• Interference Variability: Bursty interference from neighboring cells can cause rapid SINR fluctuations that outpace the CQI reporting rate, leading to link adaptation errors.

Accurate SINR estimation is the prerequisite for any link adaptation algorithm, whether classical or AI-driven.

The retransmission protocol that provides a safety net for link adaptation failures. When a transport block is decoded incorrectly, HARQ enables soft combining of multiple transmission attempts.

• Incremental Redundancy (IR): Each retransmission sends additional parity bits, progressively lowering the effective code rate until decoding succeeds.
• Chase Combining: Retransmissions are identical to the original, and the receiver combines the log-likelihood ratios (LLRs) to improve SINR.
• HARQ Processes: 5G NR supports up to 16 parallel HARQ processes to pipeline transmissions while waiting for acknowledgments.

Aggressive link adaptation that tolerates occasional first-transmission failures can maximize throughput by relying on HARQ to recover, trading off latency for spectral efficiency.

An AI-driven approach that replaces static SINR-to-MCS lookup tables with a learned policy that adapts to environment dynamics. Unlike OLLA, which is a reactive correction, DRL can learn proactive strategies.

• State Space: Includes CQI history, HARQ acknowledgment statistics, buffer status, and UE mobility estimates.
• Action Space: The selection of an MCS index for the next transmission.
• Reward Function: Typically a weighted combination of throughput (positive) and BLER penalty (negative), e.g., reward = throughput - α × BLER_violation.
• Advantage over OLLA: DRL can capture non-linear relationships and temporal dependencies, such as anticipating interference patterns from neighboring cells or adapting to UE-specific receiver capabilities.

Research using frameworks like ns-3 Gym has demonstrated DRL-based link adaptation outperforming OLLA in scenarios with high mobility and bursty interference.