A wireless receiver can only decode data if it knows how the channel is currently distorting the signal. It learns this from pilots, which are known symbols the transmitter sends in place of data. As the user moves, that estimate goes stale. Today's systems re-send pilots on a fixed schedule sized for a worst-case channel. That schedule wastes airtime when the channel is calm and drops packets when it changes faster than expected. This proposal treats the useful life of a channel estimate as a time-to-event quantity: the number of data slots until the stale estimate first causes a decoding failure. The model is learned with survival analysis, the statistics of event times when many events are never observed. Survival analysis fits because the operational data is censored by construction. Every time the base station re-trains before a failure occurs, it learns only that the estimate lasted at least that long.
The payoff is a direct, per-link answer to the classic training-throughput question. The system estimates the probability that the current estimate survives each future slot, given the pilot quality, estimated user speed, and the decoder's own confidence signals. It then picks the pilot length and re-training time that maximize expected delivered bits per channel use, subject to a per-slot failure budget. The model trains only on the acknowledgment/negative-acknowledgment (ACK/NACK) feedback the network already produces. No simulated ground-truth signal-to-noise traces are needed. It also corrects the bias that comes from the scheduler censoring its own data.
Core Claim
The claim is that a discrete-time neural hazard model yields re-training schedules with higher goodput at the same empirical reliability than analytic Doppler-based rules whenever the channel departs from the textbook fading model. Here the hazard is the per-slot probability that the next data slot fails given all previous slots succeeded. The model is trained on censored HARQ (hybrid automatic repeat request) logs and conditioned on decoder side information. Where the channel does match the textbook model, the learned schedule should roughly match the analytic rule. Three things make this more than "put a neural net on it."
First, the censoring mechanism is the scheduler itself. Treating early re-trained cycles as successes, or dropping them, biases the learned lifetime distribution. This is exactly the failure mode deep survival methods were built to avoid (Deep Learning for Survival Analysis: A Review surveys why dropping or imputing censored samples distorts the estimate).
Second, the survival curve plugs directly into a renewal-reward expression for goodput, so the decision rule needs no reinforcement learning. Let the pilot length be $L$ symbols, each data slot carry $B$ bits in $n$ symbols, and $S_j(x, L)$ be the predicted probability that the first $j$ slots after the pilot all decode given covariates $x$. Assume a NACK triggers immediate re-training. Then scheduling re-training after at most $k$ slots gives long-run goodput
$$G(L,k \mid x) = \frac{B \sum_{j=1}^{k} S_j(x,L)}{L + n \sum_{j=0}^{k-1} S_j(x,L)}, \qquad S_0 = 1,$$
which follows from the renewal-reward theorem: expected bits per cycle divided by expected cycle length. For ultra-reliable traffic we add the constraint $h_j(x,L) = 1 - S_j/S_{j-1} \le \epsilon$ for all $j \le k$, where $\epsilon$ is the tolerated per-slot failure probability.
Third, the hazard can use time-varying signals that no analytic aging model sees. Examples are decoder log-likelihood-ratio margins and iteration counts on successfully decoded slots, which tend to degrade before the first NACK.
Relation to Prior Work
The closest work is Predictive Ultra-Reliable Communication: A Survival Analysis Perspective. It models the time a scalar SNR stays above a threshold with a Weibull fit and with an MLP, then inverts the survival curve to choose a safe transmission duration. It motivates this proposal in two ways. It shows that Weibull fits break down at tail targets $\epsilon \le 10^{-2}$. It also shows that the MLP reaches an F1 of up to 0.82 at $\epsilon = 10^{-3}$ only with 10,000 fully observed training durations, and degrades sharply with 100.
It differs from our proposal on three points. Its event is an SNR crossing, not a decoding failure caused by a stale estimate. It has no pilots, so neither pilot length nor re-training time is a decision variable. It trains on fully observed simulated traces rather than on censored feedback.
The channel-aging literature supplies the model-based competitor. Sum-Rate and Power Scaling of Massive MIMO Systems with Channel Aging models aging as a first-order autoregressive process driven by the Jakes Doppler spectrum. It shows that Wiener prediction helps but has diminishing returns, and that zero-forcing receivers collapse under aged estimates far faster than maximum-ratio combining. Optimal pilot symbol power allocation under time-variant channels shows an irreducible estimation error floor at vehicular speeds up to 500 km/h. It reports roughly a 10% throughput gain from tuning the pilot-to-data power offset for least-squares estimators. Both papers constrain the proposal. They show that the aging mechanism is partly predictable from Doppler, so any learned method must beat a Doppler-aware rule, not a fixed schedule. They also show that the achievable gains from pilot tuning are on the order of ten percent, not multiples.
The finite-blocklength line determines what a slot failure means for short packets. Pilot-Assisted Short-Packet Transmission over Multiantenna Fading Channels: A 5G Case Study and A Finite-Blocklength Analysis for URLLC with Massive MIMO compute error probabilities with saddlepoint approximations of the random-coding union bound under mismatched decoding. Ferrante et al. show that shaving the pilot fraction slightly below its optimum doubles the error probability for a 30-bit payload. Error Rate Analysis for Grant-free Massive Random Access with Short-Packet Transmission makes explicit the trap that pilots raise the coding rate $B/(T-L)$. These analyses are static and offline. We reuse their error models as the simulator's ground truth for per-slot failure, and we make the pilot decision per link and over time.
The survival-methods side supplies the tools. Survival regression with proper scoring rules and monotonic neural networks proves that concordance, Brier score, and survival-CRPS are improper under censoring, while the censored log-likelihood is strictly proper. We therefore train and select models by censored log-likelihood, not F1 or concordance. Calibration and Uncertainty in Neural Time-to-Event Modeling shows that discriminative survival models are often badly calibrated in absolute terms. It also shows that calibration degrades when censoring assumptions are violated. Both problems matter here because the decision rule consumes absolute probabilities.
Research Program
The system has three parts.
The first is a slot-indexed hazard network. Its inputs are features available at the pilot: pilot length, received pilot SNR, a Doppler estimate from the phase drift between consecutive pilot estimates, the residual of the estimate, receiver type, and antenna count. At each later slot it also takes decoder side information from the slots already decoded. It outputs $h_j$ for $j$ up to a horizon. The training loss is the standard discrete-time censored likelihood. Each cycle contributes a success term for every slot it survived, plus a failure term at the NACK slot if one occurred. Cycles ended by scheduled re-training contribute no failure term, which is exactly how censoring enters.
The second part is the scheduler. It evaluates $G(L,k \mid x)$ over a small grid of pilot lengths and horizons, and it re-evaluates the remaining horizon after each slot as side information arrives.
The third part is exploration that preserves identifiability. A conservative scheduler never observes long lifetimes, so the tail is extrapolated. The logging policy therefore delays re-training by a random extra number of slots with small known probability, restricted to eMBB-class traffic where a failure is cheap. Because the policy depends only on logged covariates plus known randomization, censoring is conditionally independent of the failure time given $x$. That is the assumption the likelihood needs. We will measure how much exploration is required before tail calibration at $\epsilon = 10^{-3}$ stabilizes.
We will not rely on the fact that the hazard likelihood can be written as a per-slot binary cross-entropy. That identity is real, and it is why the training set construction, not the network, is the contribution. We state it openly and test it through the ablations below.
Baselines
The naive reference is fixed periodic re-training: a single pilot length and period per speed class, tuned by grid search on the same simulated traffic to maximize goodput subject to the reliability target. This is what a standards-configured system does. Beating it is necessary but not interesting.
The strongest competitor is a Doppler-aware analytic scheduler. It estimates the Doppler frequency from pilots and propagates the estimation error through the Jakes autoregressive aging model with Wiener prediction (Sum-Rate and Power Scaling of Massive MIMO Systems with Channel Aging). It maps the resulting effective SINR to a per-slot error probability with the saddlepoint finite-blocklength approximation (Pilot-Assisted Short-Packet Transmission over Multiantenna Fading Channels: A 5G Case Study). It then plugs these probabilities into the same goodput expression and constraint. Pairing it with our goodput rule makes the comparison isolate the probability model rather than the decision rule. It is the strongest competitor because on a channel that truly follows the Jakes model it is essentially Bayes-optimal and needs no data. I found no reported result on Lacuna for this scheduler's goodput under censored adaptive re-training, so we will build and tune it ourselves. The nearest reported numbers in the file are the roughly 10% throughput gain from optimal pilot power in Optimal pilot symbol power allocation under time-variant channels and the 0.82 F1 at $\epsilon = 10^{-3}$ in the Samarakoon et al. paper. Neither measures the same quantity. The analytic baseline gets the same pilot-length grid, a per-scenario calibration offset tuned on the validation split, and the true receiver type.
We also run the Samarakoon et al. approach: Weibull and MLP survival on SNR-threshold durations. It is given its advantage of genie SNR traces, and its predicted duration is used as the re-training period. This shows whether modeling decoding failure of a stale estimate matters relative to modeling SNR crossings.
A contextual bandit that directly selects $(L,k)$ from the same covariates and is rewarded by realized goodput tests whether the survival structure matters at all. It receives the same logged interactions, the same network capacity, and the same hyperparameter budget, measured in trials.
An oracle scheduler that knows when the true channel will cause failure gives the ceiling.
Ablations isolate each ingredient. A windowed classifier predicts failure within $k$ slots and treats censored cycles as successes. We also remove the decoder side information, remove exploration, and replace the censored log-likelihood with concordance-based model selection.
Beating the analytic scheduler means two things. The learned model must deliver higher goodput on aggregate and in every held-out scenario. It must do so while its empirical per-slot failure rate stays within the binomial confidence interval of $\epsilon$. Higher goodput bought with reliability violations counts as a loss.
Why not just…?
Why not just use the Doppler estimate and the Jakes aging model? On channels that match that model, this is probably as good as anything. We expect the learned model to tie it there, and we will report that tie as a sanity check rather than hide it. The case for learning rests on channels where the model fails. Examples are the error floor from inter-carrier interference at high speed (Optimal pilot symbol power allocation under time-variant channels), sudden blockage in ray-traced geometry, and non-stationary multipath in 3GPP urban macro channels. The zero-forcing versus maximum-ratio asymmetry (Sum-Rate and Power Scaling of Massive MIMO Systems with Channel Aging) also shows that the SINR-to-failure mapping is receiver-specific and hard to get right analytically. If the analytic rule matches the learned model on the misspecified channels too, the proposal's main claim fails, and the experiments are designed so that this outcome is visible.
Why not just train a classifier that predicts failure within the next window? Done correctly on at-risk slots with censored cycles handled properly, a per-slot classifier is the discrete hazard model. The real question is the naive version that labels scheduler-terminated cycles as successes. That version systematically underestimates late-slot hazard, because the censored cycles are exactly the ones the policy expected to be risky or long. The windowed-classifier ablation, evaluated against genie lifetimes, measures this bias directly at $\epsilon = 10^{-3}$. If the bias turns out negligible, we will say so. The contribution then reduces to the decision rule and side information.
Why not just learn the schedule end-to-end with a bandit or reinforcement learning? For throughput-only traffic this is a genuine threat, since it optimizes the target directly. It has two weaknesses. First, it gets one reward per cycle, while the hazard model extracts a label from every slot. Second, a policy learned for one $\epsilon$ must be relearned for another, whereas the survival curve serves any target by re-solving the constraint. We compare the two at matched logged-data budgets across several $\epsilon$ values. The survival approach should win on sample efficiency and on tail compliance. If the bandit matches it on both, the survival framing is not buying anything.
Proposed Experiments
Simulation covers four environments. The first is a Jakes first-order autoregressive channel, where the analytic model is correct. The second is the Vienna LTE link-level simulator at speeds up to 500 km/h, with least-squares and linear MMSE estimation. The third is QuaDRiGa 3GPP urban macro non-line-of-sight channels along mobile trajectories. The fourth is DeepMIMO outdoor scenario O1, whose ray-traced trajectories include abrupt blockage. Short-packet slots use the saddlepoint error approximation with 30-bit payloads and single-input, multiple-output (SIMO) configurations of one transmit antenna and two or four receive antennas, following Ferrante et al.
Logs are generated by the exploratory policy and split by user trajectory, so test users are unseen. Models are selected by held-out censored log-likelihood. Calibration is reported as calibration slope, following Chapfuwa et al., together with observed versus predicted hazard in each slot for the tail bins. System metrics are goodput in bits per channel use, pilot overhead fraction, and empirical per-slot failure rate with confidence intervals at $\epsilon$ of $10^{-2}$ and $10^{-3}$.
Five checks decide the outcome. The first is matched-model parity. On the Jakes channel the learned scheduler should be within noise of the analytic one. A large gain there would indicate a bug or an unfair baseline, and a large loss would indicate the learned model is data-starved.
The second is misspecification gain. On Vienna high-speed, QuaDRiGa, and DeepMIMO blockage, the learned scheduler should beat the analytic one in goodput at equal reliability.
The third is the censoring test. The naive windowed classifier and the censored hazard model are trained on logs from the same conservative policy and compared on tail calibration against genie-labeled lifetimes. This directly tests whether handling the scheduler's own censoring matters.
The fourth is sample complexity. We ask how many logged cycles it takes to meet $\epsilon = 10^{-3}$, compared against the 10,000 fully observed durations Samarakoon et al. needed, and how much exploration probability is required.
The fifth is side-information attribution. If removing decoder margins erases the gain, the honest conclusion is that decoder confidence, not survival modeling, drives the improvement, and the note's framing would be revised accordingly.
References
- Deep Learning for Survival Analysis: A Review
- Predictive Ultra-Reliable Communication: A Survival Analysis Perspective
- Sum-Rate and Power Scaling of Massive MIMO Systems With Channel Aging
- Optimal pilot symbol power allocation under time-variant channels
- Pilot-assisted short-packet transmission over multiantenna fading channels: A 5G case study
- A Finite-Blocklength Analysis for URLLC with Massive MIMO
- Error Rate Analysis for Grant-free Massive Random Access with Short-Packet Transmission
- Survival regression with proper scoring rules and monotonic neural networks
- Calibration and Uncertainty in Neural Time-to-Event Modeling