A batteryless sensor that powers itself from ambient radio or light has to decide, with very little energy, which of its options is best. A concrete case is choosing which of $K$ antenna beam directions yields the strongest link or the most harvested power. The algorithms with the strongest guarantees for this "find the best option with a fixed number of noisy trials" problem are called best-arm identification methods. In them, the chance of picking the wrong option shrinks exponentially in the number of trials. The best of these methods get that fast decay by alternating constantly between the current leader and its closest rival. On a batteryless device, alternating is expensive. Every switch reconfigures a peripheral (the antenna switch, ADC warm-up), and every brownout, when the storage capacitor runs dry and the chip loses its volatile memory, forces that reconfiguration again anyway. So the algorithm theory says to switch often, and the hardware charges a large fixed price per switch.
The claim of this note is that the tension goes away once you notice what actually controls the error rate. The exponential rate depends only on the long-run fraction of trials each option receives, not on the order in which they happen. We propose to commit to a single arm for an entire power cycle (one capacitor charge) and to set the arm for each cycle so that the running fractions track the top-two targets. This keeps the optimal error exponent while paying the switching cost once per cycle rather than once per trial. The right figure of merit then becomes error decay per harvested joule. We add a geometric prior over beam quality, precomputed offline by a Clifford-algebra wireless surrogate, which shrinks the budget needed before the exponential regime begins.
Core Claim
Let $c_m$ be the energy of one measurement, $c_s$ the energy of reconfiguring to a new arm, and $\rho^*$ the error exponent achieved by top-two sampling. Top-two sampling here follows Simple Bayesian Algorithms for Best Arm Identification, which shows that sampling the empirical leader with probability $\beta$ and a challenger otherwise achieves the optimal rate at which posterior mass on wrong arms vanishes. Under randomized top-two, a switch happens on a constant fraction $p_s$ of trials, where $p_s$ is on the order of $2\beta(1-\beta)$ once the leader stabilizes. With harvested energy $E$, the error of randomized top-two therefore decays like
$$P(\text{error}) \approx \exp\!\Big(-\frac{\rho^*}{c_m + p_s c_s}\,E\Big).$$
A commit-per-charge policy switches $S(E)=o(E)$ times, because the number of power cycles grows only as fast as energy arrives and runs are merged when consecutive cycles pick the same arm. It should therefore approach the rate $\rho^*/c_m$ per joule. The predicted gain factor is $1 + p_s c_s / c_m$, which is large exactly when peripheral reconfiguration dominates measurement cost.
The obvious objection to plain batching is that the batch sizes here are random and set by the harvester. Our theoretical task is to prove that the allocation fractions still converge to the top-two targets when run lengths are random, i.i.d.-bounded, and determined by an exogenous energy process. The proof must also show that the rate is anytime, holding at whatever moment the node is asked for an answer. The anytime property is essential because a batteryless node cannot know its total budget in advance. We build on the anytime analysis of An ε-Best-Arm Identification Algorithm for Fixed-Confidence and Beyond, whose EB-TC$_\varepsilon$ rule gives anytime exponential decay with a slack $\varepsilon$ so that nearly-equal beams do not absorb the budget. That analysis assumes unit-cost, uninterrupted sampling, and removing those assumptions is our contribution.
Where Each Direction Sits
The three directions play different roles. Exponential error decay supplies the mechanism and the guarantee. Stochastic energy harvesting sets the regime: the random batch lengths, the forced re-initialization after brownouts, and the requirement that the algorithm's state survive power loss. Geometric representation enters as a prior, not as the bandit's feature space.
The closest prior work motivates the problem but leaves it open. Pure Exploration for Constrained Best Mixed Arm Identification with a Fixed Budget obtains exponential decay under a resource constraint. That constraint, however, is a static average-cost budget, with no stochastic arrivals, no brownouts, and no cost that depends on the sequence of arms. Intermittent Inference with Nonuniformly Compressed Multi-Exit Neural Network for Energy Harvesting Powered Devices makes energy-aware runtime decisions on an MSP432 using Q-learning over capacitor voltage and output entropy. It offers no error-rate guarantee and addresses inference depth, not identification. The energy-harvesting federated learning work, Federated Learning With Energy Harvesting Devices: An MDP Framework and Decentralized Federated Learning With Energy Harvesting Devices, schedules computation against stochastic energy but targets training convergence rather than identification.
On the hardware side, two systems papers constrain the design. Graceful Performance Modulation for Power-Neutral Transient Computing Systems reports a worst-case frequency-switch overhead of 58.6 µs on the MSP430FR5739. Energy-driven computing notes that peripheral state cannot be snapshotted across brownouts, so every power cycle already pays one reconfiguration whether we switch arms or not. That observation is what makes one arm per cycle the natural batching unit rather than an arbitrary choice.
Research Program
The first component is analysis. We model energy arrivals as a stationary process driving a capacitor between turn-on and turn-off voltages. The number of trials per cycle is then a random variable $N_j$ determined by the harvest rate and $c_m$. Arm choice happens at wake-up: the policy computes the top-two leader and challenger from checkpointed statistics, then picks whichever arm's realized allocation fraction lags its target most. This is a deterministic tracking rule in place of a coin flip. We will prove convergence of fractions and the per-joule exponent under bounded-moment assumptions on $N_j$. We will also characterize the livelock boundary, the harvest rate below which one cycle cannot complete a single reconfiguration plus measurement.
The second component is state persistence. The posterior statistics are per-arm counts and sums, or a $d \times d$ ridge Gram matrix for the structured variant in the style of LinearAPT. They are a few hundred bytes, comfortably inside the 4KB RAM of the KL05Z used in ARM mbed support for transient computing. We will checkpoint them to FRAM once per cycle using a Hibernus-style snapshot. Because statistics are updated only after complete measurements, a brownout mid-measurement loses at most one sample and never corrupts the posterior.
The third component is the geometric prior. Differentiable and Learnable Wireless Simulation with Geometric Transformers (Wi-GATr) encodes scene geometry and antenna poses as projective geometric algebra multivectors, meaning points, planes, and poses packed into one 16-component object that rotates and translates consistently. It predicts received power accurately. We run it offline, at deployment or on a gateway, to predict the arm means $s_i$ for the node's $K$ beam orientations. The node's bandit then models $\mu_i = s_i + \langle \theta, x_i \rangle$, a low-dimensional linear correction on top of the surrogate. The node itself never runs a transformer; it stores $K$ numbers. The prior should shorten the long initial phase that LinearAPT reports before exponential decay appears at higher dimension.
Baselines
The naive reference is round-robin: cycle through all $K$ beams, one per power cycle, and recommend the empirical best. It switches every cycle and ignores evidence. Simple Bayesian Algorithms for Best Arm Identification shows it wastes budget on clearly inferior arms. Beating it by a large margin is necessary but not interesting.
The strongest existing algorithmic competitor is anytime EB-TC$_\varepsilon$ from An ε-Best-Arm Identification Algorithm for Fixed-Confidence and Beyond, run trial-by-trial with its adaptive leader/challenger ratio. That paper reports anytime exponential decay matching Gaussian lower bounds and outperformance of DSH, DSR, and LUCB. However, no reported result was found on Lacuna for EB-TC$_\varepsilon$ error numbers, and none for any identification method under energy harvesting. All numbers will therefore come from our own runs. It receives the same energy traces, the same measurement noise, the same $\varepsilon$, and its $\beta$ or IDS setting tuned over the same grid as ours. For structured instances we also run LinearAPT with an identical ridge regularizer and feature map.
As a learned-controller competitor we train a Q-learning scheduler in the style of Wu et al. on the same state (capacitor voltage, harvest rate, posterior entropy), with a training-trace budget equal to our tuning budget, and evaluate it on held-out traces.
The key ablation runs the identical top-two target allocation two ways, randomized per trial versus tracked per charge. It isolates the batching ingredient, and the prior-versus-no-prior ablation isolates the geometric component. Beating the baselines means a statistically clear improvement in the slope of log-error versus harvested energy on held-out traces. The improvement should grow with $c_s/c_m$ roughly as the predicted factor $1+p_s c_s/c_m$ and vanish as $c_s \to 0$. If it does not vanish, our explanation of the mechanism is wrong.
Why not just…?
Why not just add a supercapacitor or battery? Energy-driven computing argues that buffers add volume, cost, and failure modes unacceptable for miniature nodes. Hash Functions and Benchmarks for Resource Constrained Passive Devices shows that at 40 cm from an RF source, energy is too weak to charge large stores at all. More importantly, a battery removes intermittency but not the switching cost. The per-joule argument still applies with a battery, so the batching result is useful either way. We will report a continuous-power condition to show this.
Why not just run standard Thompson sampling or plain top-two and accept the switching? Russo shows plain Thompson sampling starves the runner-up and loses the optimal exponent. Plain top-two keeps the exponent per trial but pays $c_s$ on a constant fraction of trials. The randomized-versus-tracked ablation measures this directly, so this alternative is our main internal control.
Why not just use Euclidean or spherical-harmonic beam features instead of a geometric-algebra surrogate? This is a real threat to the geometric component, and we do not claim multivector features help the node's bandit itself. The claim is narrower: a better offline prior reduces the needed budget. Wi-GATr reports 0.64 dB MAE on WiPTR with 10% of training data, against 1.28 dB for PLViT and 0.69 dB for a vanilla transformer trained on all data. On real DICHASUS measurements it reports 70% lower error than a calibrated ray tracer. We will compare priors from Wi-GATr, from a ray tracer, and from no prior with degree-2 spherical-harmonic features. If the ray-tracer or feature-only variants match Wi-GATr, the geometric part of the proposal reduces to "any good prior helps," and we will say so.
Why not just learn the whole schedule with reinforcement learning? The Q-learning baseline tests this empirically. Its drawbacks are that it offers no anytime error guarantee and must be retrained per harvester and per deployment. We expect it to be competitive in-distribution and to degrade on held-out trace types.
Proposed Experiments
Arm means come from WiPTR scenes (12,000 indoor layouts) and Wi3R layouts. For each scene we place a receiver and define $K \in \{8, 32, 128\}$ beam orientations. Ground-truth received power comes from ray tracing, measurements are Gaussian around it, and the Wi-GATr prediction serves as the prior. Energy arrivals are replayed from the Southampton harvester trace repository (indoor PV, RFID, kinetic; DOI 10.5258/SOTON/404058) through a capacitor model. We use $c_m$ and $c_s$ measured on an MSP430FR5969 and sweep $c_s/c_m$ over two orders of magnitude.
The decisive checks are four. First, the log-error versus harvested-energy slope of tracked allocation should approach $\rho^*/c_m$, while randomized top-two sits at the degraded slope. Second, the gap between them should track $1+p_s c_s/c_m$ across the sweep. Third, the Wi-GATr prior should shorten the energy required before the linear-slope regime begins, relative to the ray-tracer prior and to no prior. Fourth, no policy should livelock above the predicted harvest-rate threshold.
A final hardware run puts the policy on an MSP430FR5969 under RF harvesting, as in Su et al., with a switched-beam antenna. We report misidentification rate, switches per joule, and checkpoint energy, and validate the sim-to-real gap using DICHASUS-derived arm means. An exploratory extension lets the chosen beam change the node's own harvest, so exploring bad beams also shrinks the budget. We will report whether energy-weighted target fractions improve the per-joule rate there.
References
- Simple Bayesian Algorithms for Best Arm Identification
- An $\varepsilon$-Best-Arm Identification Algorithm for Fixed-Confidence and Beyond
- Pure Exploration for Constrained Best Mixed Arm Identification with a Fixed Budget
- Intermittent Inference with Nonuniformly Compressed Multi-Exit Neural Network for Energy Harvesting Powered Devices
- Federated Learning With Energy Harvesting Devices: An MDP Framework
- Decentralized Federated Learning With Energy Harvesting Devices
- Graceful Performance Modulation for Power-Neutral Transient Computing Systems
- Energy-driven computing: Rethinking the design of energy harvesting systems
- LinearAPT: An Adaptive Algorithm for the Fixed-Budget Thresholding Linear Bandit Problem
- ARM mbed support for transient computing in energy harvesting IoT systems
- Differentiable and Learnable Wireless Simulation with Geometric Transformers
- Hash Functions and Benchmarks for Resource Constrained Passive Devices: A Preliminary Study