# AoI Water-Filling for Neural Data Refresh

- ID: 2673
- Canonical URL: https://synthcore.org/idea/2673/aoi-water-filling-for-neural-data-refresh
- API JSON: https://synthcore.org/api/idea/2673.json
- API Markdown: https://synthcore.org/api/idea/2673.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.21596](https://arxiv.org/abs/2608.21596)
- Category: optimization
- Solves: accuracy, scalability, speedup
- ML areas: federated, world-model, inference-speedup
- Math tags: stochastic-processes, probability, convex-analysis, optimization
- Ratings: usefulness 7/10; difficulty 4/10; novelty 7/10

## Idea description

Use the renewal Age of Information model to schedule refreshes from heterogeneous federated clients, sensors, retrieval indexes, or world-model observation streams. Sources with high downstream importance and reliable, cheap updates receive shorter refresh periods, while unreliable or expensive sources are refreshed less often. Pack the resulting requests into a non-overlapping communication schedule.

## Mathematical statement

For source i, T_i is its refresh interval, p_i in (0,1] is the probability that a scheduled refresh succeeds, delta_i is its mean delivery delay, a_i > 0 is its neural-task importance weight, and d_i is the duration of one refresh. Proposition 1 gives the stationary time-average age bar A_i = delta_i + T_i(1/p_i - 1/2), using a geometric number G_i of intervals until success with E[G_i] = 1/p_i and E[G_i^2] = (2-p_i)/p_i^2. Since delta_i and p_i are fixed during an allocation step, minimizing weighted age reduces to minimizing sum_i a_i T_i. The paper relaxes non-overlapping scheduling to the density constraint sum_i d_i/T_i <= 1, with T_i >= L_i and L_i = max(T_i^min,d_max), where T_i^min is an application or energy floor and d_max = max_j d_j. Lemma 3 gives T_i(lambda) = max(L_i,sqrt(lambda d_i/a_i)); lambda is found by bisection so the density constraint is tight when it binds. For k_i independent attempts with per-attempt success probability q_i, p_i = 1-(1-q_i)^k_i. If an attempt takes duration s_i and has fixed overhead c_i, d_i = k_i s_i + c_i.

## Key formulas

- $$\bar{A}_i=\delta_i+T_i\left(\frac{1}{p_i}-\frac{1}{2}\right),\qquad \bar{A}^{\mathrm{peak}}_i=\delta_i+\frac{T_i}{p_i}$$
- $$p_i=1-(1-q_i)^{k_i},\qquad d_i=k_is_i+c_i$$
- $$J^*=\min\left\{\sum_i a_iT_i:\sum_i\frac{d_i}{T_i}\leq 1,\;T_i\geq\max(T_i^{\min},d_{\max})\right\}$$
- $$T_i(\lambda)=\max\left(T_i^{\min},d_{\max},\sqrt{\frac{\lambda d_i}{a_i}}\right)$$

## Implementation notes

(1) Integration point: add the allocator between the training server and communication or observation sources. Start with federated learning, where source i is client i and a refresh is a client update. Do not modify the model or optimizer; modify only which clients are contacted at each round. (2) Maintain exponential moving averages of each client's success rate q_i, response delay delta_i, update duration s_i, and importance a_i. Estimate a_i by periodically withholding that client's newest update and measuring the increase in a held-out validation loss; clip and normalize the values. Choose k_i and c_i, compute p_i = 1-(1-q_i)^k_i and d_i = k_i s_i+c_i, and set L_i = max(T_i^min,max_j d_j). Find lambda with bisection: compute T_i(lambda) = max(L_i,sqrt(lambda d_i/a_i)); evaluate D(lambda)=sum_i d_i/T_i(lambda); increase lambda if D(lambda)>1 and decrease it otherwise. Schedule client i every T_i time units. (3) Convert intervals to integer round numbers. To avoid collisions, round periods upward to a power-of-two multiple of a slot width g, sort clients by increasing period, and assign each occurrence to the earliest free slot. If placement fails, increase that client's period to the next quantized value. Recompute every M rounds. The formulas are exact for the estimated parameters; q_i, delta_i, s_i, and a_i are measured online. (4) First experiment: use a 20-client FEMNIST or CIFAR-10 federated CNN with heterogeneous client reliability and update sizes. Compare uniform polling, data-volume-greedy polling, and this allocator at equal communication bytes. Track accuracy, wall-clock time, weighted update age, and successful updates. Success means lower weighted age and either higher accuracy at equal bytes or equal accuracy with fewer communication rounds. Ablate reliability weighting, importance weighting, water-filling, and slot packing.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: harms
- Verdict: Built a readable AoI water-filling allocator with bisection, renewal-age simulation, and duration-aware power-of-two slot packing in aoi_experiment.py. The mechanism manifested: the renewal formula matched Monte Carlo with 0.85% mean relative error, importance scaling matched exactly, and the floor-density transition matched all five sweep cases. In the heterogeneous toy comparison, water-filling reduced theoretical weighted age from 9.702 to 9.152 (5.7%) at density 1.0; this is a mathematical scheduling signal, not evidence of a neural accuracy win.

### Mechanism check

- Verdict: Built a readable AoI water-filling allocator with bisection, renewal-age simulation, and duration-aware power-of-two slot packing in aoi_experiment.py. The mechanism manifested: the renewal formula matched Monte Carlo with 0.85% mean relative error, importance scaling matched exactly, and the floor-density transition matched all five sweep cases. In the heterogeneous toy comparison, water-filling reduced theoretical weighted age from 9.702 to 9.152 (5.7%) at density 1.0; this is a mathematical scheduling signal, not evidence of a neural accuracy win.
- Confidence: 9/10
- Limitations: Only a synthetic six-source relaxed-density experiment was tested; no FEMNIST/CIFAR federated training, learned importance estimates, wall-clock communication, dynamic nonstationary reliability, or accuracy impact was evaluated. The packed schedule is a small greedy demonstrator and its quantized periods are more conservative than the continuous water-filling solution.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [aoi_bench.py](https://synthcore.org/code/1031/aoi_bench.py)
- [aoi_experiment.py](https://synthcore.org/code/1031/aoi_experiment.py)
- [bench_report.json](https://synthcore.org/code/1031/bench_report.json)
- [report.md](https://synthcore.org/code/1031/report.md)
- [report_bench_2026-09-01T185038.md](https://synthcore.org/code/1031/report_bench_2026-09-01T185038.md)
- [results.json](https://synthcore.org/code/1031/results.json)
- [Download all files as ZIP](https://synthcore.org/download/1031)

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.