- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper studies neural network training under varying renewable energy supply. It models training segments as renewal frames with action-dependent durations and uses a virtual queue to constrain error averaged over training time. A phase-dependent weight changes the queue penalty, and a calibrated predictor scores candidate batch sizes before measured feedback updates the state. The paper presents long-run guarantees and evaluates shortfall and target attainment using image classification training and a replayed photovoltaic trace.
Critical strengths
- Jointly considering supply timing, variable action duration, and learning progress is a meaningful problem.
- The evaluation executes closed-loop GPU training rather than reporting predicted trajectories alone. Shared-checkpoint calibration and measured feedback are sensible design choices.
- The paper reports common-quality-target metrics, full-run metrics, and unfavorable midnight results.
- Small-workload scope, time-compressed PV replay, slack constraints, and overlapping phase-weight outcomes are explicitly acknowledged.
Critical weaknesses
- The central ratio guarantee has a counterexample. Equation 7 minimizes a separate conditional ratio after observing each state, whereas Theorem 4.1 compares a ratio of expected totals. Let independent states A/B each occur with probability one half. All errors are zero and sigma=1, keeping the queue zero with strict feasibility. Actions a/b have (shortfall,duration)=(0,1)/(1,100) in A and identical (100,1) outcomes in B. The stated rule selects a in A, yielding long-run rate 50. The fixed policy always selecting b yields rate 1. This satisfies bounded costs and positive bounded durations, and the nonvanishing gap contradicts the claimed O(1/V) result. The comparator need not even depend on the state. Global fractional optimization requires a shared ratio parameter in the additive objective; independently minimizing state-specific ratios is different. Appendix J does not justify that substitution.
- The Markov proof has an additional distribution mismatch. Equation 89 substitutes stationary costs under another policy for a current-state one-step comparison. Uniform mixing of each fixed policy does not make an adaptive policy occupy the comparator's distribution or incorporate future state values. Policy-dependent dynamics generally require a suitable bias/occupancy or multistep argument, not an appended vanishing term.
- The phase-dependent proof omits the switching term. Equation 37 freezes the current phase, but the actual Lyapunov increment includes (w_next-w_current)Q_next²/2. Such frozen-phase increments do not telescope along changing phases. Positive switches are possible under the stated detector, and their quadratic contribution is not bounded by a fixed U_max. Theorem 5.1 consequently lacks its central justification. The finite-state Markov description also omits the queue on which action selection depends.
- Average predictor RMSE does not establish the uniform per-frame C-additive oracle accuracy assumed by the theory, particularly when errors are multiplied by backlog and phase weights.
- All experimental error constraints are slack, on one small MNIST workload. Thus the constraint/backlog tradeoff is untested. Reported training time excludes validation/controller costs despite subsecond target times, and calibration cost also matters. Compressed PV replay supports a trace experiment, not a deployment over a physical day.
- Frame-indexed independent supply is not the same external wall-clock trace across policies with different durations. Endpoint-loss weighting is also a defined proxy rather than the actual within-frame loss integral.
Questions for authors
Q1. How do Theorem 4.1 and Algorithm 1 address the two-state counterexample in W1, where the stated rule has rate 50 but fixed action b has rate 1? Which precise assumption excludes it, or must the controller instead use a shared global ratio parameter?
Q2. What justifies replacing a current-state one-step comparison with another policy's stationary costs in Appendix L, Equation 89? Where are differing occupancy distributions and future-state values handled under controlled transitions?
Q3. How does the telescoping argument account for (w_next-w_current)Q_next²/2? What bounds or assumptions control this term when phases can switch in both directions?
Q4. Since decisions depend on Q, why is the process containing only training state and phase a finite time-homogeneous Markov chain? Must the state or policy assumptions change?
Q5. How does average prediction RMSE imply the required per-frame C-additive accuracy when errors are multiplied by Q or w_phi Q? If it does not, which guarantees apply to the implemented controller?
Q6. Do the energy and target-time advantages persist with binding error budgets, a common wall-clock supply trace, and validation/controller/calibration costs included?