Wireless sensor nodes that share one radio channel can teach themselves a collision-free schedule. Each node repeatedly tries time slots in a repeating frame and keeps the slot where its packets get acknowledged. This works in simulation and breaks on real radios. When a hardware acknowledgment is lost, the node cannot tell whether its slot was taken by a newcomer or whether its own transceiver hiccuped, so it either abandons a good slot or clings to a bad one. The learner also needs to know the right frame length in advance, which means knowing how many nodes exist. We propose to fix both problems with one signal. The coordinating sink runs a small recursive estimator of how many nodes are contending. It broadcasts in each beacon how surprised that estimator was by the last frame. Nodes use the broadcast to set how fast they forget. When the network as a whole looks stationary, a local failure is treated as noise and old evidence is kept. When the network looks like it changed, nodes forget quickly and re-explore. The sink uses the same estimate to grow or shrink the frame.
The per-node learning rule is simple enough to run as spike-timing-dependent plasticity (STDP), a local learning rule in spiking neurons that strengthens a connection when an input spike precedes an output spike. That makes the scheme a candidate for event-driven, microwatt-scale hardware. The spiking implementation is not the core claim, though. The core claim is that a global change signal, computed where the whole channel is visible, is what lets local learners separate their own noise from real network change. We test the spiking version only as a faithful realization of an explicit Bayesian rule.
Core Claim
The closest existing system is ALOHA-Q-DEPS from Use of Q-learning approaches for practical medium access control in wireless sensor networks. In that protocol, each node runs Q-learning (a trial-and-error value update) over slots in a fixed frame. The authors found that on IRIS hardware plain ALOHA-Q loses a slot after 7 consecutive acknowledgment failures. They patched this with a heuristic based on long-run success rate. They also state that the frame size must be derived from prior topology knowledge. A 12-node grid needed 23 slots in practice versus 20 in simulation.
On the estimation side, From Static to Dynamic Tag Population Estimation: An Extended Kalman Filter Perspective tracks a time-varying RFID tag population from idle, success and collision counts. It uses an extended Kalman filter, a recursive estimator that linearizes a nonlinear model at each step. It adds CUSUM, a running change-detection statistic on the filter's prediction errors, to decide when to adapt. Its analysis relies on large-population approximations (10^4 to 10^5 tags), and it never feeds its change signal back to the devices.
Our claim is that broadcasting the sink's filter surprise as a forgetting gate for each node's slot beliefs does three things at once. First, it removes the dependence on prior frame sizing. Second, it makes learned schedules immune to local acknowledgment loss during stationary periods. Third, it lets schedules reconverge after nodes join or leave faster than any single fixed forgetting rate can. Neither prior line combines a network-level change detector with distributed slot learners, and neither handles frames whose length changes while reservations must survive.
Research Program
The sink maintains an exact recursive Bayesian posterior over the number of active contenders $N$ on a discrete grid. We use a grid rather than an extended Kalman filter because at 8–21 nodes the large-system approximations of Yu et al. do not hold, and a grid of a few dozen values is cheap. After each frame the sink computes the surprise $s_k = -\log p(o_k \mid o_{1:k-1})$, where $o_k$ is the vector of idle, success and collision counts in frame $k$. It passes this surprise through a calibrated squashing function, a CUSUM-style accumulator whose threshold is set from the surprise distribution under the model's own stationary predictions rather than hand-tuned. The result is a gate $g_k \in [0,1]$ carried in the next beacon.
Each node keeps discounted success and failure counts for every slot:
$$\alpha_s \leftarrow \lambda_k \alpha_s + a_s, \qquad \beta_s \leftarrow \lambda_k \beta_s + (1-a_s), \qquad \lambda_k = 1 - g_k(1-\lambda_{\min}).$$
Here $a_s$ is 1 if the node transmitted in slot $s$ and was acknowledged, and 0 if it transmitted and was not. Slots the node did not use are left unchanged. $\lambda_{\min}$ is the fastest allowed forgetting rate. Each node picks its slot by Thompson sampling, meaning it draws a success probability for every slot from its Beta$(\alpha_s,\beta_s)$ belief and uses the best draw. When $g_k \approx 0$, memory is effectively unbounded. This is the "infinite memory when stationary" behavior that Yu et al. obtain from CUSUM, and it is the property that should neutralize isolated acknowledgment drops.
Frame length is set by the sink. Slots that carried repeated successes count as reserved. The sink appends slots for the estimated number of unreserved contenders, following the classical frame-slotted ALOHA rule of roughly one slot per contender that underlies the dynamic framed protocols of Medium Access Control Protocols for Wireless Sensor Networks with Energy Harvesting. The frame grows by appending and shrinks by trimming unreserved tail slots, so existing reservations keep their slot indices. Handling reservation-preserving resizing is a design problem the fixed-frame learners never faced, and we will analyze it explicitly.
The spiking realization maps a node's slots to a winner-take-all pool of neurons, meaning the neurons compete and only one fires. The winning neuron's spike triggers a transmission, and an acknowledgment arriving inside a timing window acts as the post-event that potentiates that neuron's weight. STDP enables spiking neurons to detect hidden causes of their inputs shows that such circuits perform online expectation–maximization, with weights settling at log-probabilities. That result supplies the method and gives the reason to expect the weights to approximate $\log$ success probability. The gate $g_k$ enters as a multiplicative learning-rate modulator.
Two findings constrain the spiking design. Spike timing-dependent plasticity as dynamic filter shows that pair-based rules mispredict multi-spike bursts and that suppression terms are needed to avoid runaway potentiation, so we include its suppression variable. The spiking message-passing work of A Spiking Neural Network Implementation of Gaussian Belief Propagation and Spike-Timing-Dependent Plasticity for Bernoulli Message Passing needed population codes or a supervised teacher layer. We avoid both, because each node here learns only a per-slot Bernoulli quantity, and the acknowledgment itself is the naturally occurring teaching signal. A second loose analogy supports the gate. Gated KalmaNet: A Fading Memory Layer Through Test-Time Ridge Regression shows that input-dependent forgetting in a Kalman-style recurrence beats uniform decay in sequence models. We move that gating decision from the learner to an external observer that sees the whole channel.
Baselines
The naive reference is fixed-frame slotted ALOHA with random slot choice and a frame length computed from the true node count. This gives an oracle-sized frame with no learning. Every learner must beat its throughput.
The strongest existing method is ALOHA-Q-DEPS Use of Q-learning approaches for practical medium access control in wireless sensor networks, run with its oracle frame size and its historical-success heuristic. Its reported ceiling performance is 0.48 Erlangs of throughput on the 21-node random topology, against a theoretical 0.5. The file reports no result for DEPS under node churn or swept acknowledgment-loss rates. We reimplement it in the same simulator, on the same topologies and loss models, with the same tuning budget over its exploration constants. Because DEPS is given the true frame size, matching its 0.48-Erlang stationary throughput without topology knowledge counts as parity. Beating it means strictly lower reconvergence time after churn and strictly fewer spurious slot abandonments at equal throughput.
A third baseline is a centralized reservation scheme. The sink estimates $N$ with the same grid filter and explicitly assigns slots via join requests, with the beacon overhead charged in slots. We also run a "sink-sized DEPS" hybrid, which pairs our frame sizing with DEPS learners. It isolates whether the gain comes from sizing or from gated forgetting.
The key-ingredient ablations replace $g_k$ in three ways. The first is a constant $\lambda$, grid-searched per scenario with the best value reported. The second is a per-node local CUSUM on the node's own acknowledgment stream. The third is a gate delayed by several frames or randomly shuffled across frames. If a well-tuned constant $\lambda$ matches the gated rule across both the loss sweep and the churn sweep, the core claim fails.
Why not just…?
Why not just use a per-node Beta posterior with a well-chosen fixed discount? This is the most direct threat. It also predicts the classic trade-off Yu et al. document for fixed forgetting. A slow discount survives acknowledgment loss but reconverges slowly after churn, and a fast discount does the reverse. The best-tuned constant-$\lambda$ ablation settles this directly. We will report the full Pareto curve of spurious abandonments against reconvergence time and check whether the gated rule lies outside it.
Why not just run change detection locally on each node? A node sees only its own slot. A newcomer colliding with it and its own transceiver dropping acknowledgments look identical from inside one slot. That is exactly the failure behind the 7-drop collapse in Use of Q-learning approaches for practical medium access control in wireless sensor networks. The sink sees idle and collision counts across the whole frame, and local hardware loss does not move those counts. The local-CUSUM ablation measures how much this separation is worth. If local detection matches the gated rule, the broadcast bit is unnecessary and we will say so.
Why not just let the sink assign slots centrally? If the sink can broadcast, explicit assignment is a real competitor, and in small static networks it may win. Its costs are per-node addressing in beacons, which grow with $N$, a join handshake that itself contends for the channel, and the fact that the sink cannot tell which nodes collided. We charge those costs in the centralized baseline and report the network size and churn rate at which each approach wins, rather than claiming dominance.
Why not just skip spiking hardware and run the Beta rule on a microcontroller? The protocol stands on the Beta rule alone, and that version is the primary artifact. The spiking version is evaluated only on whether its slot choices and throughput match the Beta rule within simulation noise. If they diverge, the neuromorphic realization is reported as not yet faithful, and the protocol result is unaffected.
Proposed Experiments
We reproduce the three topologies used in Use of Q-learning approaches for practical medium access control in wireless sensor networks in a slot-level simulator: an 8-node linear chain, a 12-node grid and a 21-node random deployment. We add a hardware acknowledgment-loss model that is independent per node and swept from 0 to 30%. We calibrate it so that plain ALOHA-Q shows the reported collapse after 7 consecutive drops, and we use a TinyOS/IRIS testbed if one is available.
Three decisive checks follow. The first is a stationary acknowledgment-loss sweep. The prediction is near-zero spurious slot abandonments for the gated rule at loss rates where greedy ALOHA-Q collapses and constant-$\lambda$ learners either abandon slots or fail after churn. The second is a churn sweep in which nodes join and leave on Poisson schedules. The prediction is reconvergence time below the best constant-$\lambda$ and below DEPS, with frame length tracking the true contender count with no topology prior. The third is the shuffled-gate control. If shuffling the gate does not hurt, the gain is not coming from change detection.
Metrics are throughput in Erlangs, reconvergence frames, spurious abandonments per thousand frames, idle-slot listening energy, and sink tracking error of $N$. As an out-of-regime stress test for the sink filter alone, we run it on the dense RFID setting of Capture-Aware Dense Tag Identification Using RFID Systems in Vehicular Networks, which reports 1.94% cardinality error at 1000 tags. We compare against that figure only for the cardinality estimate, since capture effects are outside our node model.
References
- Use of Q-learning approaches for practical medium access control in wireless sensor networks
- From Static to Dynamic Tag Population Estimation: An Extended Kalman Filter Perspective
- Medium Access Control Protocols for Wireless Sensor Networks with Energy Harvesting
- STDP enables spiking neurons to detect hidden causes of their inputs
- Spike timing-dependent plasticity as dynamic filter
- A Spiking Neural Network Implementation of Gaussian Belief Propagation
- Spike-Timing-Dependent Plasticity for Bernoulli Message Passing
- Gated KalmaNet: A Fading Memory Layer Through Test-Time Ridge Regression
- Capture-Aware Dense Tag Identification Using RFID Systems in Vehicular Networks