Censoring-Aware Trust Regions for Survival-Constrained Optimization

When a learned failure model is placed inside an optimizer, the optimizer will push toward wherever the model looks most optimistic. Usually that is a region where the model has never seen data. Constraint-learning methods guard against this by restricting decisions to the convex hull of the training points, meaning the smallest region "spanned" by the observed examples. With right-censored data that guard leaks. Right-censored records are records where observation stopped before the event, so we only know the asset lasted at least that long. Suppose a battery run under an aggressive charging regime was watched for only two months before observation stopped. Its record says it survived at least two months. It says nothing about whether it survives the twelve-month planning horizon. Yet a standard hull counts that regime as fully supported. If aggressive regimes are also the newest, and so the most heavily censored, the optimizer will tend to pick exactly the regimes whose long-run survival nobody has observed.

We propose a simple fix and a careful test of whether it matters. The trust region is indexed by the planning horizon $H$. It is built only from units whose record actually informs survival to $H$: those still under observation at $H$, plus those observed to fail. Units censored before $H$ are excluded. The construction stays representable as linear constraints with integer variables, so it slots into a mixed-integer linear program (MILP). A MILP is an optimization model mixing continuous decisions such as power levels with on/off decisions, and it is solved to certified optimality by branch-and-bound. We embed a survival model trained on censored data into a storage-dispatch MILP and then ask a concrete question: does the censoring-aware region reduce how often the chosen schedule truly violates its survival target? We measure that rate on testbeds where the true lifetimes are known.

Core Claim

Let unit $i$ have operating covariates $x_i$, observed time $y_i = \min(T_i, C_i)$ (the earlier of failure time $T_i$ and censoring time $C_i$), and event indicator $\delta_i$, which equals 1 if failure was observed. We define the horizon-$H$ support set and trust region as

$$ \mathcal{D}_H = \{\, i : y_i \ge H \ \text{or}\ \delta_i = 1 \,\}, \qquad \mathcal{X}_H = \operatorname{conv}\{x_i : i \in \mathcal{D}_H\}. $$

The decision $x$ must satisfy the learned chance constraint $\hat S(H \mid x) \ge 1-\varepsilon$, where $\hat S$ is the model's probability of surviving past $H$ and $\varepsilon$ is the tolerated failure risk. It must also lie in $\mathcal{X}_H$, which is encoded with weights $\lambda_i \ge 0$, $\sum_i \lambda_i = 1$, $x = \sum_i \lambda_i x_i$, and $\lambda_i = 0$ for $i \notin \mathcal{D}_H$.

The claim is not that this is mathematically deep. The claim is that it addresses a failure mode that model-side fixes cannot. Censoring that depends on the operating regime breaks positivity: beyond the follow-up window there is simply no data on survival to $H$. Reweighting, censored likelihoods, and ensembles all correct bias within the observed follow-up window. None of them creates evidence beyond it. We predict that the gap between the standard hull and $\mathcal{X}_H$ will grow with the censoring rate and with the correlation between censoring and regime aggressiveness. We also predict that the gap disappears under random censoring. That last prediction is the key falsifiable check.

Relation to Prior Work

The method directly extends Mixed-Integer Optimization with Constraint Learning (OptiCL). OptiCL embeds trees, ensembles and multi-layer perceptrons into mixed-integer programs, and it uses a convex-hull trust region together with an ensemble wrapper. It even treats a chemotherapy problem with a survival outcome. Its trust region, however, treats every training row as equally informative, which is exactly the assumption that censoring breaks. We reuse its embedding machinery unchanged and replace only the region.

Two papers supply the survival models we embed. Optimal Survival Trees: A Dynamic Programming Approach (SurTree) and Optimal Sparse Survival Trees (OSST) produce globally optimal survival trees by dynamic programming but stop at offline prediction. A tree's leaves translate directly into MILP routing constraints, as in Tight mixed-integer optimization formulations for prescriptive trees. We also embed a discrete-time hazard model in the spirit of Limitations of Binary Classification for Long-Horizon Diagnosis Prediction and Advantages of a Discrete-Time Time-to-Event Approach: Empirical Analysis and of multi-task logistic regression from Learning Patient-Specific Cancer Survival Distributions as a Sequence of Dependent Regressors. For a piecewise-exponential model with log-hazard linear in $x$, the constraint $-\log \hat S(H\mid x) = \sum_k \Delta_k e^{\beta_k^\top x} \le -\log(1-\varepsilon)$ is convex in $x$. Here $\Delta_k$ is the length of time bin $k$ and $\beta_k$ is that bin's coefficient vector. Because the constraint is convex, it can be outer-approximated by tangent cuts, which keeps the problem a MILP.

Several papers motivate the work. The dispatch literature models wear deterministically. Storage Degradation Aware Economic Dispatch uses a Rainflow-based quadratic depth-of-discharge cost. Health-aware energy management for multiple stack hydrogen fuel cell and battery hybrid systems uses fixed start/stop and load-change penalties and reports a 64.68% cost reduction against collective stack control. By contrast, fleet studies show that real battery failure is a covariate-driven, heavily censored process. Energy-Based Survival Models for Predictive Maintenance reports 74% censoring and Variable selection for heavy-duty vehicle battery failure prognostics using random survival forests reports over 90%.

Two papers constrain the design. The binary-classification analysis above found that dropping early-censored subjects does not fix bias. That is why we exclude censored units from the region while keeping them in the likelihood: they still carry valid lower-bound information about survival before $C_i$. Effective Ways to Build and Evaluate Individual Survival Distributions shows that the concordance index (C-index), a ranking measure of how well predicted risks order observed failures, is insufficient. We therefore also require D-calibration of embedded models, a test of whether predicted survival probabilities are calibrated across the whole time axis, so that $\hat S(H\mid x)$ is meaningful as a probability.

Research Program

Survival models. We build three embeddable survival models: an OSST/SurTree survival tree, a piecewise-exponential hazard model, and a small discrete-time neural hazard model embedded through OptiCL's MLP encoding. Each is trained with a censored likelihood, or with an Integrated Brier Score objective for OSST, on the same data.

Decision problem. The decision is a battery operating envelope: caps on charge rate and depth of discharge, and a temperature setpoint. This envelope bounds an hourly arbitrage dispatch over the price profiles used in the Storage Degradation Aware Economic Dispatch paper. The envelope must keep $\hat S(H\mid x) \ge 1-\varepsilon$.

Trust regions. We compare four regions. The first is the standard OptiCL hull over all units. The second is $\mathcal{X}_H$. The third is a hull over failed units only. The fourth is a tree-native variant that forbids leaves whose longest follow-up is shorter than $H$. That last variant is the cheapest version of the same idea, and we want to know whether it suffices.

Theory. Analytically, we will state the positivity argument formally. We will also show that under censoring independent of $x$, $\mathcal{X}_H$ and the full hull coincide in the large-sample limit. This turns the random-censoring control into a sharp prediction.

Baselines

Naive reference. The do-nothing reference is the deterministic-proxy dispatch from Storage Degradation Aware Economic Dispatch. It uses a Rainflow cycle cost with no survival constraint, the same price data, and the same solver. It shows how much any survival constraint changes decisions and true failure rates. Beating it means a lower true violation rate at comparable revenue.

Strongest competitor. The strongest existing method is OptiCL with its standard convex-hull trust region and ensemble wrapper, embedding the identical censored-likelihood survival model. This is the right competitor because it already does everything except treat censored rows differently. Any gain we report is then attributable to the region alone. The comparison will use the same trained models, the same ensemble size, the same solver time limit, and the same $\varepsilon$. Beating it means a statistically lower true constraint-violation rate at matched $\varepsilon$, with revenue loss smaller than the violation-rate reduction justifies. We also report regret against an oracle MILP trained on uncensored data. There is no reported result found on Lacuna for censored-data decision quality of OptiCL or any comparable method. The nearest reported numbers concern prediction, not decisions: CyclePatch reaches a mean absolute percentage error (MAPE) of 0.179 on lithium-ion cycle life in BatteryLife: A Comprehensive Dataset and Benchmark for Battery Life Prediction, which serves only as a reference for how predictable lifetimes are with full data.

Ablations. The failed-units-only hull isolates whether censored units carry useful region information. The leaf-follow-up rule isolates whether a tree-native version captures most of the benefit. A model-side ensemble lower confidence bound with no hull tests whether uncertainty alone does the job.

Why not just…?

Why not just reweight the training data for censoring and keep the standard hull? Inverse-probability-of-censoring weighting removes bias when censoring depends on covariates, but only where some units remain under observation at $H$. Where follow-up never reaches $H$, the weights are undefined, and that is exactly the region the optimizer exploits. We include an IPCW-trained model inside the standard hull as an additional arm. The prediction is that it reduces violations under mild regime-dependent censoring and fails under strong regime-dependent censoring.

Why not just use an ensemble's pessimistic survival estimate as the constraint? This is the real threat. If ensemble disagreement widens beyond the follow-up window, the hull may be redundant. Our concern is that ensemble members share an inductive bias, such as smooth log-linear hazards or flat tree leaves, and so they agree confidently when extrapolating. The ensemble-lower-bound arm settles this directly. If it matches $\mathcal{X}_H$ on violation rate at similar revenue, we will report the hull as unnecessary for that model class.

Why not just binarize at the horizon and embed a classifier? Limitations of Binary Classification for Long-Horizon Diagnosis Prediction and Advantages of a Discrete-Time Time-to-Event Approach: Empirical Analysis reports a time-dependent AUC of 0.47 for the best binary classifier against 0.72 for a discrete-time model on ADHD at age 7. Dropping censored-before-$\tau$ units also discards up to 90% of fleet data. We include this as a weak arm for completeness.

Why not just keep deterministic wear penalties? They contain no lower-bound information from surviving assets and cannot express a calibrated failure probability. The naive-reference arm measures what that costs in true failures.

Proposed Experiments

Testbed A: real lifetimes with imposed censoring. BatteryLife's run-to-failure cells span 990 batteries, 59 chemistries and 9 temperatures. Because every cell has a known lifetime, we can impose synthetic right-censoring at 50%, 74% and 90%, matching the fleet rates above. We vary one knob: how strongly censoring time shortens with charge-rate aggressiveness, from independent censoring to strongly regime-dependent censoring. Ground-truth survival of a chosen envelope comes from a model fit to the full uncensored data and is cross-checked against the empirical lifetimes of the nearest real cells.

Testbed B: exact ground truth. A Weibull simulator follows the synthetic battery setup of Energy-Based Survival Models for Predictive Maintenance, so true $S(H\mid x)$ is exact.

Testbed C: tractability. The fuel-cell/battery bus problem on the Chinese Bus Driving Cycle (CBDC) from the health-aware energy management paper tests solve-time overhead in a larger mixed-integer model. In that problem we replace the deterministic stack-wear penalties with an embedded survival constraint.

Metrics. We report true violation rate of the survival target, realized revenue or cost, regret against the uncensored oracle, and solve time. As embedding prerequisites, we also report C-index, Integrated Brier Score, and D-calibration of each embedded model.

Decisive checks. The proposal succeeds if three things hold. First, under regime-dependent censoring, $\mathcal{X}_H$ cuts true violations relative to the OptiCL hull at every censoring level, with the gap growing from 50% to 90%. Second, under independent censoring the two regions perform indistinguishably. Third, the ensemble-lower-bound arm does not close the gap. If the second check fails, the effect is not about censoring. If the third fails, the cheaper model-side fix wins, and we will say so.