PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Channel Hopping for IEEE 802.15.4 TSCH-Based Networks: A Dynamic Bernoulli Bandit Approach

Nastooh Taheri Javan, Masoud Sabaei, Vesal Hakami

arXiv:2609.27876v1cs.NI

TL;DR

Fixed and single-channel IEEE 802.15.4 hopping can be ill-suited to changing external interference, while existing blacklisting methods assume stationary interference statistics. The paper models adaptive TSCH hopping as a dynamic bandit and proposes a lightweight learner that tracks changes; simulations report improved PDR and energy efficiency and lower end-to-end delay.

  • Problem

    Existing adaptive blacklisting schemes assume stationary external-interference statistics, although interference can change over time.

    Source: §I · p. 2

  • Method

    The paper formulates adaptive channel hopping as a dynamic MAB problem and uses a lightweight learning algorithm to track time-varying interference for channel selection.

    Source: §I · pp. 2–3

  • Results

    Accounting for non-stationary interference improves PDR and energy efficiency and lowers end-to-end delay in the reported simulations.

    Source: §VIII · pp. 13–14

  • Takeaways & Limitations

    The algorithm’s low complexity and demonstrated memory footprint fit the stated constraints of embedded IoT devices.

    Source: §VIII · p. 13

Abstract

from arXiv · show

In IEEE 802.15.4 standard for low-power low-range wireless communications, only one channel is employed for transmission which can result in increased energy consumption, high network delay and poor packet delivery ratio (PDR). In the subsequent IEEE 802.15.4-2015 standard, a Time-slotted Channel Hopping (TSCH) mechanism has been developed which allows for a periodic yet fixed frequency hopping pattern over 16 different channels. Unfortunately, however, most of these channels are susceptible to high-power coexisting Wi-Fi signal interference and to possibly some other ISM-band transmissions. This interference manifests itself in the form of the presence/absence of other devices with either or both static and dynamic channel selection policies. In order to isolate channels with undesirable conditions, blacklisting mechanisms are defined to adapt the channel hopping process. However, the existing solutions which form blacklists unrealistically assume that the statistical model of the external interference remains fixed, and do not vary over time. In this paper, we realistically assume that the impact of external interferes on 802.15.4 may generally follow a non-stationary pattern, and accordingly formulate the adaptive channel hopping problem as a Dynamic Multi-Armed Bernoulli Bandit (Dynamic MABB) process from the machine learning theory. We then propose an online learning algorithm with track-ability properties for computing an adaptive hopping policy. Simulations confirm that when the statistics of the external interference has a switching regime, the proposed solution outperforms the previous schemes in terms of both energy efficiency as well as two important KPIs for TSCH-based networks, i.e., PDR and latency.

I. INTRODUCTION

Fixed channel hopping can perform poorly under interference, while existing adaptive blacklisting approaches assume interference statistics remain stationary. This paper frames adaptive TSCH hopping as a dynamic bandit problem and proposes a lightweight learning approach for non-stationary conditions.

I. INTRODUCTION

  • Default TSCH uses a fixed periodic hopping pattern across 16 channels, many of which can be affected by Wi-Fi and other ISM-band transmissions.Single-channel IEEE 802.15.4 communications can also face interference and fading associated with high delay and low packet delivery ratio.Source: §I · p. 1
  • Existing blacklisting schemes assume time-invariant interference statistics, although real external users and channel-selection behavior can change dynamically.The paper argues that this makes fixed statistical models a poor fit for real-world spectrum use.Source: §I · p. 2
  • The authors model channel selection as a Dynamic Multi-Armed Bernoulli Bandit, with channels as arms and packet transmission success or failure as Bernoulli outcomes.The non-stationary formulation reflects interference whose statistical behavior changes over time.Source: §I · p. 2
  • They deploy AFF-OTS, an adaptive-filtering-inspired learner that tracks evolving data by gradually reducing the weight of older observations.The procedure is described as lightweight, robust to tuning parameters, and initialized without prior knowledge of model structure.Source: §I · p. 2
  • Adaptive hopping requires multi-offset scheduling because default single-offset scheduling offers each link only one channel per time-slot.The paper uses a graph-theoretic multi-offset scheduling scheme so nodes can select an offset through the learning procedure.Source: §I · p. 2
  • The proposed scheme is evaluated in simulations against default TSCH and stationary-interference work, with a physical setup used to assess memory footprint.The stated objectives include maximizing average PDR and tracking time-varying interference with low-cost recursive updates.Source: §I · pp. 2–3

II. BACKGROUND

The background contrasts TSCH’s uniform, blind channel hopping with adaptive methods that learn channel quality. It reviews prior schemes and their trade-offs in sensing cost, coordination, and assumptions.

A. Default channel hopping in TSCH

  • Default TSCH maps each time-slot and scheduled channel offset to a physical frequency, uniformly selecting channels without regard to their quality.The mapping uses the network’s absolute sequence number, the current offset, and a bijection over available channels.Source: A. Default channel hopping in TSCH · p. 3

B. Adaptive Channel Hopping

  • Adaptive hopping seeks to avoid low-PDR channels, while model-free learners estimate channel quality from experience without requiring a known channel-dynamics model.The background notes that such schemes are relevant when channel conditions change over time or cannot be reliably characterized before deployment.Source: B. Adaptive Channel Hopping · p. 3

C. Related Work

  • A-TSCH selectively uses reliable channels but relies on costly spectrum sensing; ETSCH [16] senses during idle slots, with timing and multi-hop challenges.ETSCH+DCS combines central and distributed estimation, but adds end-node overhead and assumes coordinator reachability to all nodes.Source: C. Related Work · pp. 3–4
  • Other reviewed methods use transmitter-side channel lists [20] or negotiated per-link blacklists, [21], with constraints including rate knowledge, scheduling conflicts, and neighborhood size.MABO-TSCH [21] is reported to deliver 23% more throughput than basic frequency hopping and select the best channels in 75% of cases.Source: C. Related Work · p. 4
  • Related work also includes temporary blacklists requiring network-wide distribution and experiments on airplane-cabin interference reporting about 35% packet error rate.The reviewed reinforcement-learning experiments evaluate PDR under simulated wireless interference and multipath fading.Source: C. Related Work · p. 4

D. Motivation

The paper motivates adaptive channel hopping by noting that external interference changes over time, while existing TSCH blacklisting methods assume stationary interference. It focuses on external interference in a tree-like network and introduces a lightweight Dynamic MABB algorithm for dynamic hopping.

D. Motivation

  • Existing adaptive TSCH blacklisting schemes assume stationary interference, although external sources can appear, disappear, or change channel use over time.The paper contrasts this assumption with the dynamic behavior of real-world interference.Source: D. Motivation · p. 4
  • The proposed approach applies Dynamic MABB through a lightweight algorithm that IoT nodes iteratively use to determine their hopping pattern.The paper presents the method as addressing non-stationary external interference in TSCH networks.Source: D. Motivation · p. 4

III. SYSTEM MODEL

  • The system model is a tree-like network where nodes send data upstream to a gateway.Useful links run from child to parent nodes.Source: §III · p. 5, Figure 1 · p. 5

A. Interference Model

  • The model distinguishes internal interference, addressed by scheduling, from external interference caused by other equipment using the IEEE 802.15.4 frequency band.It focuses on external interference and assumes a collision fully corrupts the useful signal.Source: A. Interference Model · p. 5
  • The interference model represents TSCH and interferer channel offsets as 𝕁 and 𝕀, with γ_ij indicating whether an offset pair interferes.The occurrence probability is p_ij, and absence probability is 1 − p_ij.Source: A. Interference Model · p. 5
  • Under the stationary-interference assumption, the probability of interference on each channel offset remains constant over time.The model represents interference occurrence or absence on offset j as Y_j(n).Source: A. Interference Model · p. 5

2) Non-stationary external interference

Non-stationary interference is described through a sequence of probabilistic regimes for each channel offset. The paper makes no general assumption about regime switching, though its simulations emphasize Markovian switching.

2) Non-stationary external interference

  • For each channel offset, Ψ_j is a finite, countable set of probabilistic regimes, each associated with a distribution, finite average, and finite variance.The active regime may change over time.Source: 2) Non-stationary external interference · p. 5
  • The regime governing an offset at time n determines its corresponding probability distribution.The text denotes the active regime as ψ_j(n) and its distribution as F_ψ_j(n)(.).Source: 2) Non-stationary external interference · p. 5
  • The approach assumes no specific regime-switching process; simulations emphasize the special case of switching according to a first-order Markov chain.In that case, regimes are states of a finite-state Markov chain with transitions specified by a stochastic matrix.Source: 2) Non-stationary external interference · p. 5

B. Framing Model and Multi-Offset Scheduling

Adaptive hopping is paired with a multi-offset TSCH schedule and formulated as a non-stationary Bernoulli bandit problem. The objective is to maximize expected packet delivery reward, with low regret requiring the learner to track changes in the optimal channel.

B. Framing Model and Multi-Offset Scheduling

  • The multi-offset scheduler uses a conflict graph and two phases to allocate time-slots and channel offsets, assigning unused offsets to maximize network throughput.It accounts for nonallowable simultaneous transmissions and hidden-terminal interference.Source: B. Framing Model and Multi-Offset… · p. 6
  • The hopping algorithm selects one channel from the legitimate offsets available to a link in its scheduled time-slot.The channel-slot schedule repeats, and links eventually use different physical channels through the cyclic slot-frame pattern.Source: B. Framing Model and Multi-Offset… · pp. 5–6, Figure 2 · p. 6
  • The proposed hopping scheme is orthogonal to the scheduler if it provides multiple offsets per link and time-slot, and can run with centralized or distributed scheduling.The algorithm takes allocated slot-frame time-slots and associated channel offsets as inputs.Source: B. Framing Model and Multi-Offset… · p. 6

A. Background on multi-armed bandits

  • In the bandit formulation, channels are arms, each device is an agent, and expected packet delivery ratio is the objective under time-varying reward probabilities.The changing interference can shift which channel is optimal, so the paper adopts Dynamic MABB for Bernoulli transmission outcomes.Source: A. Background on multi-armed bandits · p. 6

B. Problem formulation

  • At each scheduled time-slot, a receiving node chooses a channel; the objective is to maximize total expected reward or equivalently minimize regret.The formulation treats successful or failed transmissions as binary outcomes.Source: B. Problem formulation · p. 6
  • The proposed learning algorithm seeks sub-linear regret by swiftly identifying switches in the optimal channel, so the expected reward gap vanishes in the long run.The comparator is a policy that selects the optimal channel at each moment.Source: B. Problem formulation · p. 7

V. THE PROPOSED ADAPTIVE HOPPING ALGORITHM

The algorithm adapts AFF-OTS, a dynamic MABB solver, to channel hopping by tracking channel PDR with adaptive forgetting and selecting channels from evolving reward beliefs.

V. THE PROPOSED ADAPTIVE HOPPING ALGORITHM

  • AFF-OTS is chosen for fewer computational steps and more flexible tuning, and its simulations in report lower total regret than competing schemes.The cited schemes are described as automating key-parameter tuning without prior knowledge of the model structure, which is useful for resource-limited IoT devices.Source: §V · p. 7
  • The dynamic MABB procedure separates learning each channel’s reward distribution from selecting a channel for transmission.In this setting, the learned reward distribution corresponds to channel PDR; the paper leaves technical details to.Source: §V · p. 7

A. Estimating the expected PDR of each channel

  • The estimator computes a sample mean of transmission outcomes while progressively reducing the weight of older feedback through adaptive forgetting.Its forgetting factor can change at each time step to respond to evolving interference without prior knowledge of that process.Source: A. Estimating the expected PDR of each… · p. 7
  • The forgetting factor is adjusted with a gradient-descent step based on an online derivative-like prediction-error function, then constrained to [0,1].The derivative-like function uses precomputed values, and the step size is described as much smaller than 1.Source: A. Estimating the expected PDR of each… · pp. 7–8

B. Updating PDR estimates for unselected channels

Because a node observes only its selected channel, the algorithm discounts stored quantities for unselected channels so their PDR estimates continue to update.

B. Updating PDR estimates for unselected channels

  • Discounting the stored quantities for unselected channels keeps their PDR estimates updating, addressing the loss of tracking precision when new observations stop.The quantities are discounted using the forgetting factor from the channel’s last selection; when selected, the discounted and regular quantities match.Source: B. Updating PDR estimates for… · p. 8

C. Adaptive channel selection scheme

DMABB-CH selects channels using Bayesian reward beliefs, with optimistic sampling and discounted estimates that increase exploration of channels left unselected.

C. Adaptive channel selection scheme

  • For each channel, Thompson Sampling draws from its Beta posterior and selects the channel with the highest sample.The Bernoulli reward model uses a Beta conjugate prior whose posterior parameters are updated recursively.Source: C. Adaptive channel selection scheme · p. 8
  • Discounting unselected channels’ quantities flattens their posterior distributions, increasingly boosting their exploration the longer they remain unselected.The algorithm applies Optimistic Thompson Sampling, replacing a sampled score with the posterior mean when that mean is greater.Source: C. Adaptive channel selection scheme · pp. 8–9
  • Initialization normally lasts O(|J|), selecting each channel once before the main algorithm loop.The receiver initially follows the default TSCH hopping sequence across the physical channels.Source: C. Adaptive channel selection scheme · pp. 8–9
  • Each receiver runs channel selection independently for each child link, records reception success as reward 1 and failure as 0, and piggybacks its next-channel choice in feedback.The paper describes the feedback overhead as negligible.Source: C. Adaptive channel selection scheme · p. 9

D. Discussion on computational complexity

Each iteration updates channel beliefs and samples candidate scores, using basic operations that the authors characterize as low processing overhead.

D. Discussion on computational complexity

  • For the typically 16 IEEE 802.15.4 channels, each iteration performs per-channel random sampling, comparison and recursive updates, then maximizes over channel scores.The parameter updates use primitive addition, subtraction and multiplication operations.Source: D. Discussion on computational… · p. 9
  • A parent executes DMABB-CH separately for each child, but only one instance runs on a node per timeslot, yielding low processing overhead and a reasonably small memory footprint.The memory-footprint characterization assumes a tree topology with a moderate branching factor.Source: D. Discussion on computational… · p. 9

VI. SIMULATION RESULTS

The simulations use tree-topology networks with stationary and time-varying Wi-Fi interference, comparing DMABB-CH against default TSCH hopping and a static MAB-based scheme.

A. Simulation setup

  • Simulations use randomly placed tree-network nodes in a 200 × 200 m² area, with 20–40 nodes and 50 m coverage radius.NS-3 is used, and Wi-Fi is the modeled external interferer.Source: A. Simulation setup · p. 9
  • The comparison includes default periodic TSCH hopping and a static MAB-based method alongside the proposed approach.Default TSCH uses a single-offset schedule; the MAB baseline uses multi-offset scheduling but is formulated for stationary interference.Source: A. Simulation setup · p. 10
  • Interference scenarios include a stationary model and slow or fast Markov switching among four probability matrices.The slow transition matrix is near-diagonal, extending expected time between regime changes; the fast scenario uses a separate transition matrix.Source: A. Simulation setup · p. 10

B. Simulation Results

Across the simulated settings, DMABB-CH adapts to changing interference and is reported to improve PDR, energy efficiency, and end-to-end latency relative to the baselines.

Convergence

  • In the fast-switching scenario, normalized regret for DMABB-CH approaches zero as the time horizon grows.The authors identify fast switching as the most challenging case; other cases show less noisy waveforms and relatively faster convergence.Source: B. Simulation Results · p. 10, Figure 3 · p. 10

PDR

  • Under stationary interference, MAB-based hopping has nearly the same limiting PDR as DMABB-CH, but performs progressively worse as interference dynamics increase; default TSCH fares worst.Environmental dynamics degrade PDR, with the largest impact on default TSCH, which does not adapt.Source: B. Simulation Results · p. 11, Figure 4 · p. 11
  • PDR decreases as network size grows with fixed scheduling resources and as offered load intensifies.The network-size and offered-load comparisons are shown in Figs. 5 and 6, respectively.Source: B. Simulation Results · p. 11, Figure 5 · p. 11, Figure 6 · p. 11
  • For stochastic interference with memory, higher-order models improve MAB-based PDR, while standard TSCH shows no meaningful relationship with memory level.The paper reports these comparisons in Table VII.Source: B. Simulation Results · p. 11, Table VII · p. 11

Energy efficiency

  • Across network sizes, DMABB-CH improves energy efficiency over both baselines by reducing retransmission attempts.Energy efficiency is defined as energy required per successfully delivered packet or bit.Source: B. Simulation Results · p. 11, Figure 7 · p. 11

Average end-to-end delay

  • Across network sizes, DMABB-CH yields fewer retransmissions and lower end-to-end latency than the compared schemes.End-to-end delay trends slightly upward as network size increases.Source: B. Simulation Results · p. 12, Figure 8 · p. 12

VII. EXPERIMENTAL SETUP

The experimental setup measures DMABB-CH's embedded memory footprint and compares physical-testbed measurements with simulations under similar conditions.

A. Implementation setup and memory footprint

B. Performance measurements

The physical-testbed comparison uses PAR and single-hop delay to assess channel-hopping performance, while the conclusion reports improved PDR, energy efficiency, and end-to-end delay when interference non-stationarity is accounted for.

B. Performance measurements

VIII. CONCLUSION

  • Accounting for interference non-stationarity improved PDR, with higher energy efficiency and lower end-to-end delay in the reported results.Source: §VIII · pp. 13–14
Loading 2609.27876v1…