Pilots on a Survival Clock: Scheduling Channel Re-training from Censored CSI Lifetimes

A wireless receiver can only decode data if it knows how the radio channel is currently distorting the signal. This knowledge, called channel state information (CSI), is learned from pilot symbols: known sequences that occupy airtime that could otherwise carry data. Today pilots are mostly sent on a fixed schedule. When a user moves quickly, the CSI goes stale before the next pilot arrives and packets are lost. When the channel barely changes, the schedule wastes capacity. This note proposes treating "how long will the current channel estimate stay good enough to decode?" as a time-to-event question. Survival analysis is the branch of statistics built for such questions. It predicts the full distribution of when an event will happen, and it correctly handles observations that end before the event occurs, which are called censored observations. The receiver would predict a survival curve for the CSI it has just estimated, and from that curve it would choose when to send the next pilot so as to maximize net throughput.

This is a natural fit because of where the training data comes from. The receiver never sees its true CSI error. It learns that stale CSI caused a failure only when a block fails its checksum, and every observation stops the moment a new pilot refreshes the estimate. CSI lifetimes are therefore right-censored by construction, and the censoring is caused by the receiver's own pilot decisions. Plain regression on observed lifetimes is biased in exactly this setting. Survival likelihoods are designed for it.

Core Claim

A deep survival model trained with a censoring-aware likelihood can produce calibrated survival curves for CSI validity, conditioned on what the receiver sees at each pilot: the channel estimate, its recent change, and a Doppler estimate. Plugged into a simple renewal-reward rule, these curves should achieve higher net spectral efficiency than fixed pilot periods, Doppler-derived periods, and pilot triggers driven by channel predictors. The gain should be largest in channels whose decorrelation is not captured by a single Doppler number, such as 3GPP geometry-based channels where scattering clusters appear and disappear.

The scheduling rule is where the survival curve earns its place. Suppose a pilot costs $T_p$ symbols, the next pilot is sent $\tau$ symbols later, and data received after the CSI fails at time $T$ is lost. By the renewal-reward theorem, which says long-run throughput equals expected reward per cycle divided by expected cycle length, the long-run fraction of useful symbols is

$$ \eta(\tau \mid x) = \frac{\int_0^{\tau} S(t \mid x)\, dt}{T_p + \tau}, $$

Here $S(t \mid x)$ is the probability that the CSI estimated from pilot features $x$ is still valid $t$ symbols later. Setting the derivative to zero, the optimal interval satisfies

$$ S(\tau^* \mid x)\,(T_p + \tau^*) = \int_0^{\tau^*} S(t \mid x)\, dt . $$

The rule needs the whole calibrated curve, in absolute terms. A risk ranking alone, which is what a Cox model's partial likelihood identifies, is not enough. Neither is a point prediction of lifetime. This requirement is what separates the proposal from simpler predictors, and it is why the training objective matters.

Relation to Prior Work

The closest work is Predictive Ultra-Reliable Communication: A Survival Analysis Perspective. It models how long a link stays above a signal-strength threshold, using both a Weibull fit and an MLP, and derives a maximum transmission duration for a target reliability. That paper motivates ours. It also shows the parametric model drifting when the fading departs from Weibull, while the neural model reaches an F1 of 0.82 at $\epsilon=10^{-3}$ with 10,000 samples. Its limitations are the ones we address. The event is defined on received power rather than on estimation error. There are no pilots in the model. Lifetimes are fully observed in simulation, so censoring is never confronted. The only channel is synthetic Rayleigh flat fading.

Asymptotically optimal pilot allocation over Markovian fading channels schedules pilots across users with a Whittle index, a per-user priority score for restless bandit problems. It shows that pilot priority should jump once CSI age passes a threshold. That threshold structure is consistent with our rule. However, the paper assumes a small discrete Markov chain and perfect CSI after every pilot. We replace the chain with a learned, covariate-dependent survival curve and use noisy pilot estimates.

On the survival side, the method comes from Survival regression with proper scoring rules and monotonic neural networks. SuMo-net enforces a monotone survival function, trains on the exact right-censored log-likelihood, and argues that concordance and Brier-type objectives are improper. Improper here means they can favor a wrong distribution, which would directly corrupt $\tau^*$. Its fast inference also matters, because the model must run once per pilot epoch on baseband hardware. The latency concern raised by MambaCSP applies here too. As a second model family we use discrete-time conditional-hazard networks of the kind described in Neural Networks to Predict Survival from RNA-seq Data in Oncology, since checksum outcomes arrive per block. The review Deep Learning for Survival Analysis: A Review acts as a constraint. Its point that Cox losses are biased under minibatch risk sets is one more reason to prefer likelihood-based models that output absolute curves. Its observation that competing risks are understudied shapes one of our analyses.

Data-aided estimators such as Semi-Data-Aided Channel Estimation for MIMO Systems via Reinforcement Learning and Detect to Learn attack overhead from a different direction. They reuse decoded data as extra pilots within a frame. Our approach is complementary: it decides when real pilots are needed at all. We include a data-aided receiver as a robustness condition rather than as a competitor.

Research Program

The first step is to define an event that a deployed receiver can actually observe. In simulation we have the true channel, so we record two event definitions. The oracle event is the first symbol at which the error of the held channel estimate exceeds a threshold $\gamma$. The deployable event is the first block whose checksum fails while a fresh-CSI decoder would have succeeded. Every lifetime is right-censored at the next pilot or at the end of the session. Checksum failures caused by noise rather than staleness form a competing risk. We will test whether modeling them as a separate cause, rather than folding them into censoring, changes $\tau^*$. This goes directly at the competing-risks gap flagged in the review.

The second step is the model itself. Covariates at each pilot are the current channel estimate, compressed to its dominant spatial and delay components, the difference from the previous estimate, an estimated Doppler spread, SNR, and the time since the previous pilot. A SuMo-net and a discrete-hazard network are both trained with the right-censored likelihood.

The third step addresses a feedback problem. Once the model sets the schedule, it decides where censoring falls, and it will rarely observe lifetimes longer than its own predictions. We will randomize pilot timing with small jitter and record the probability of each chosen interval. We then compare naive likelihood training against inverse-propensity-weighted training and measure the bias that arises when the system collects its own data.

Baselines

The naive reference is a fixed pilot period of the kind used in current frame structures. We sweep the period and report the best single period chosen in hindsight on the test distribution. This is generous to the baseline, and beating it shows that adaptivity matters, not just a better constant.

The strongest physics baseline is a Doppler rule. The receiver estimates the maximum Doppler shift and sets the pilot period to a fixed fraction of the implied coherence time, with the fraction tuned on validation data. It is cheap, standard, and close to optimal under classical isotropic scattering, so it is the baseline most likely to tie us.

The strongest learned baseline is a channel-prediction trigger. A Kalman/autoregressive predictor and a MambaCSP-style sequence model (MambaCSP) forecast the channel, and a pilot is sent when their predicted error growth crosses a threshold tuned on validation data. MambaCSP reports 9% lower NMSE in TDD and 12% lower in FDD than LLM-based predictors on QuaDRiGa UMa NLOS at 10–100 km/h. That is the best reported number in this file for channel prediction on our benchmark, though it is a prediction-accuracy figure and not a scheduling result.

The prior survival baseline is a Weibull MLE curve plugged into the same renewal rule, following Predictive Ultra-Reliable Communication. In the multi-user extension we add the Whittle index from Asymptotically optimal pilot allocation over Markovian fading channels, run on a channel discretized into its required number of states.

For the end-to-end metric, net spectral efficiency under adaptive pilot scheduling on QuaDRiGa, no reported result was found on Lacuna, so every number will be produced in-house. All methods get identical channel realizations, receiver chains, and covariate access. Learned methods get the same training set and the same hyperparameter search budget.

Two ablations isolate the key ingredient. The first trains the same architecture with mean squared error on observed lifetimes, once treating censored lifetimes as events and once dropping them. The second removes schedule randomization. Beating these baselines means higher net spectral efficiency than the best of the Doppler rule and the predictor trigger at a matched or lower block-error rate. The win must hold across speeds, not only on average.

Why not just…?

Why not just set the pilot period from the estimated Doppler spread? Under Clarke-type fading with a single Doppler parameter, this rule may be nearly optimal, and we expect it to tie on the synthetic Rayleigh benchmark. The case for survival modeling rests on geometry-based channels. Tracking Time-Varying Channels Beyond the Quasi-Static Assumption notes that cluster birth and death and non-stationarity within a frame break single-parameter descriptions. If the Doppler rule matches us on QuaDRiGa as well, the honest conclusion is that learned curves are unnecessary, and we will report that.

Why not just predict the channel and send a pilot when predicted error grows? A predictor does not know its own error at deployment time. Its error is revealed only at pilots or checksum failures, which is censored data again. Calibrating a threshold on the predictor amounts to an implicit, uncalibrated survival model. The predictor-trigger baseline, with its threshold tuned on validation data, tests this directly. Feeding the predictor's output into our survival model as a covariate tests whether the two are complementary.

Why not just fit a Weibull curve, or a Cox model? The Weibull model is the right choice when data is scarce, and Predictive Ultra-Reliable Communication reports that it is accurate at moderate reliability with about 100 samples. We therefore report the crossover in training-set size. Cox models only identify relative risk, but $\tau^*$ needs the absolute $S(t \mid x)$. Using Cox would also mean estimating a separate baseline hazard and accepting minibatch risk-set bias (Deep Learning for Survival Analysis: A Review).

Proposed Experiments

The main benchmark is QuaDRiGa 3GPP UMa/UMi NLOS with 4×4 and 64-antenna arrays at 10–100 km/h, matching the settings in Detect to Learn and MambaCSP. We include mixed-speed sessions in which a user accelerates mid-session. We also reproduce the synthetic time-correlated Rayleigh URC benchmark from Predictive Ultra-Reliable Communication, which gives a direct comparison of a censoring-aware likelihood against its MLP and Weibull models. As a mmWave stress test we use DeepMIMO trajectories. Metrics are net spectral efficiency, block-error rate, pilot fraction, and the right-censored log-likelihood. We also report calibration of $S(t \mid x)$ at the predicted $\tau^*$.

Three checks decide the outcome. First, on QuaDRiGa the survival policy must beat the tuned Doppler rule and the predictor trigger in net spectral efficiency at equal block-error rate. Second, the censoring-naive ablations must produce measurably miscalibrated curves and worse $\tau^*$. If they do not, the survival machinery is decoration. Third, without randomization the self-collected training loop should drift toward overly short lifetimes and pilot over-provisioning, and propensity weighting should correct this.