Autonomous QoE — the BNG learns each subscriber's normal, and fixes the ones that drift
loss · RTT · QoE
the quiet degradations
relaxes when they recover
it ever changes a packet
1 · Why a fixed threshold is the wrong tool
Every QoE alert today is built on a number an operator typed in: loss > X%, latency > Y ms. That number is a compromise that fits no one. Set it tight and a noisy fixed-wireless cell pages you all night; set it loose and a fibre customer whose line quietly slid from 0.2% to 1.5% loss — a fourfold degradation they can absolutely feel on a video call — sails straight under it. The threshold can't tell the difference between a line that's always a bit lossy and a line that just got worse, because it has no idea what “normal” means for that specific subscriber.
2 · The idea — learn the normal, act on the deviation
Autonomous QoE is a control loop that sits on top of the Flow Intelligence engine. Flow Intelligence already measures, per subscriber and passively in the XDP data plane, the real signals — download loss, access vs transit RTT, and a 0–100 experience score. Autonomous QoE turns those measurements into a self-driving loop: learn → detect → act → relax.
3 · The three pillars
- A rolling EWMA baseline per subscriber for loss%, access RTT, transit RTT and QoE score — plus the variance of each.
- Learned entirely from the signals Flow Intelligence already measures — no new probes, no per-subscriber configuration.
- Cold-start safe: a brand-new subscriber is judged against a fleet-wide baseline until it has enough of its own history, so it never throws false alarms on day one.
- Anomaly = how many standard deviations a subscriber's current loss / RTT / QoE sits from its own learned baseline.
- Catches the quiet degradations a fixed threshold misses — the fibre line that slid 0.2% → 1.5%, invisible to a 5% rule but obvious against its own history.
- Each signal is judged only when it was actually measured, so a line with no latency sample this moment never skews its own baseline.
- A subscriber that stays anomalous for several consecutive intervals (hysteresis — never one noisy tick) earns a graduated push: its bulk/elephant flows are demoted harder so its interactive traffic recovers.
- The push is scaled by how far off normal it is, and decays the moment it recovers — back to baseline, no human action.
- It only ever raises aggressiveness above the operator baseline; it never drops below your floor, and it reverts on reload.
4 · What it helps
Stop tuning thresholds
Caught before the call
No human in the loop
5 · How it works — under the hood
The whole loop is a userspace control layer in the daemon — it reads the per-subscriber signals Flow Intelligence already computes and writes a single per-subscriber demotion hint. It adds no new data-plane code and no new fast-path cost: the measurement was already happening; this is the intelligence on top.
Baseline & anomaly
For every subscriber the daemon keeps a slow exponentially-weighted moving average and variance of each metric. The anomaly score is the positive, degradation-direction z-score — (current − baseline) / stddev — taking the worst across loss, access RTT, transit RTT and QoE. A subscriber is “anomalous” when that exceeds a small multiple of its own deviation and it has learned enough history; before then it is measured against a fleet baseline so fresh lines stay quiet. Each metric is fed only when it carried a real sample, so an unmeasured signal can never teach the baseline a false “normal.”
Hysteresis & the autonomous push
Detection alone would flap. So a subscriber must be anomalous for several consecutive intervals before it earns a push, and the push is a graduated demotion boost scaled by how far off normal it is. That boost folds into the same per-subscriber remediation the loss engine already drives — one decision per subscriber, taking whichever signal says “lean harder” — so the two never fight. The instant the subscriber's streak breaks, the boost is withdrawn and the line decays back to its operator baseline.
mode : enforce
warmed baselines : 1377
anomalous (now) : 14
enforced (boost) : 2
— most-anomalous —
ppp967 z=62 txRTT 510ms (base 18) QoE 0 (base 97)One glance: the box has learned 1,377 baselines, 14 subscribers are worse than their own normal right now, and 2 have drifted long enough to be acted on — here, a line whose transit latency jumped 28× over its own baseline.
6 · Safety & deployment
| Stage | What it does | Risk |
|---|---|---|
| off | nothing — no baselines kept | none |
| observe | learns every subscriber's baseline, flags anomalies, recommends — never changes a packet | none — read-only |
| enforce | sustained-anomalous subscribers get a graduated, decaying demotion boost on their bulk flows, layered on the operator baseline | bounded — hysteresis + ceiling + auto-revert |
Stop tuning thresholds. Let the BNG learn each subscriber's normal — and fix the ones that drift.
Autonomous QoE turns Flow Intelligence's measurements into a self-driving loop: it learns every subscriber's own baseline, detects who just got worse than they normally are, and leans on exactly those lines — with hysteresis, automatic decay, and a floor you set. No probes. No per-subscriber config. No dashboard to watch. Observe-first, default-off.
Turn on observe today, read who the box says is drifting, and enable enforce when the recommendations earn your trust.
fi baseline snippet is illustrative of the output format, not a benchmark; no specific performance figures are claimed. Related per-topic briefs (Flow Intelligence, L4S/AQM, Interactive Flow Protection, QoS, NOC2) are available alongside this guide.