Probabilistic Node Selection (FedPNS)

• Probabilistic Node Selection (FedPNS) is a framework that models client participation as continuous probabilities to replace deterministic node scheduling in federated learning.
• It leverages explicit probabilistic optimization and alternating algorithms for joint node selection and power allocation to balance energy, delay, and fairness.
• Empirical evaluations highlight 3–5× speedup, significant energy savings, and improved fairness in non-IID and asynchronous network environments.

Probabilistic Node Selection (FedPNS) refers to a family of methodologies that replace deterministic client or node scheduling in distributed and federated machine learning with explicit, analytically-tractable probability models for node selection. This strategy circumvents the computational intractability and lack of performance guarantees inherent to mixed-integer deterministic schemes, and provides powerful mechanisms to balance fairness, convergence, energy efficiency, privacy, and other operational constraints that are otherwise difficult to encode or optimize in classical approaches. The FedPNS paradigm has been formalized in a range of settings, including wireless federated learning, asynchronous and decentralized topologies, non-IID data regimes, and highly heterogeneous GPU/cloud environments (Marnissi et al., 2024, Wu et al., 2021, Yang et al., 2023, Stanisic et al., 11 Nov 2025, Zec et al., 2023).

1. Fundamental Principles and System Models

FedPNS frameworks start by generalizing binary (selected/not selected) participation to continuous selection probabilities $p_i \in (0, 1]$ for each node or client. In each round, participation is determined by an independent Bernoulli trial of bias $p_i$. This construction enables optimization of expected participation or related objectives, subject to operational constraints. Key model elements include:

• System Topology: Typically involves one or more servers and a large set of wireless devices or edge clients, potentially with heterogeneity in bandwidth, channel quality, data distribution, energy budget, and computational resources (Marnissi et al., 2024, Yang et al., 2023, Stanisic et al., 11 Nov 2025).
• Costs and Constraints: Device computation (CPU/GPU cycles), transmit power, communication energy, upload/download bandwidth, and per-round energy/time budgets are explicitly parameterized. For each client $i$, costs may include $E_i(P_i)$ (energy per round, as a function of power), $T_i(P_i)$ (transmission delay), and SLOs (service-level objectives) such as deadlines or minimum probability of timely completion.
• Participatory Fairness: Nonzero probability floor $p_i \geq \lambda > 0$ can be imposed to guarantee participation diversity, mitigate starvation/straggler effects, and improve learning for rare or underrepresented data (Yang et al., 2023).

2. Optimization Formulations and Solution Algorithms

The cornerstone of FedPNS is an explicit probabilistic optimization framework, allowing for tractable solutions via continuous relaxations. Representative problem classes include:

• Joint Probabilistic Selection and Power Allocation: Maximize the expected weighted number of participating clients, where weights $w_i$ reflect dataset size or importance, subject to per-client energy ($p_i\,E_i(P_i) \leq E_i^\text{max}$) and time ($p_i\,T_i(P_i) \leq T_{\text{th}}$) constraints, and hardware limits ($0 \leq P_i \leq P_\text{max}$). The full problem is

$p_i$0

(Marnissi et al., 2024).

• Alternating Optimization: Non-convexity due to variable coupling is handled via block coordinate-ascent or alternating minimization. Fixing one block (e.g., $p_i$1), the other (e.g., $p_i$2) is solved optimally, sometimes with closed form via fractional programming (e.g., Dinkelbach's method).
• Gradient-Driven Probability Updates: In settings where convergence speed is paramount (e.g., non-IID federated learning), node selection probabilities are adapted based on gradient alignment metrics. Suboptimal or "adverse" clients—whose gradients are anti-aligned with the current global update—have their $p_i$3 decremented via an explicit update rule, and the participation mass is redistributed proportionally among better-aligned nodes. This yields faster convergence and diminishes the impact of poorly aligned local updates (Wu et al., 2021).
• Bandwidth and SLO-aware Selection: For GPU-heterogeneous, space-ground federated topologies, selection is based on analytic forecasting of round-trip latency and service-level probabilities (e.g., probability $p_i$4 that client $p_i$5 completes within deadline $p_i$6). This supports selection in highly dynamic environments without continuous monitoring (Stanisic et al., 11 Nov 2025).

3. Analytical Insights and Theoretical Guarantees

FedPNS approaches offer several analytical advantages over deterministic heuristics or mixed-integer scheduling:

• Convexity After Relaxation: With continuous probabilities, objectives become convex or quasi-concave in relevant variables, enabling global or locally optimal closed-form updates (Marnissi et al., 2024).
• Elimination of Mixed-Integer Programming: Binary constraints are relaxed, entirely avoiding the computational complexity and heuristic solutions required for deterministic selection.
• Convergence Guarantees: Algorithms monotonically increase (or leave unchanged) the expected objective in each iteration, and objective upper bounds (e.g., $p_i$7) ensure convergence to a stationary point (Marnissi et al., 2024, Wu et al., 2021).
• Fairness and Inclusiveness: Nonzero probabilities for all nodes prevent permanent exclusion of high-cost or rare-data nodes, supporting fair and robust aggregation (Yang et al., 2023).
• Trade-off Control: Explicit parameters allow for tunable balances between convergence rate, energy consumption, straggler mitigation, SLO compliance, and communication fairness. For example, in (Yang et al., 2023), parameter $p_i$8 controls prioritization of convergence vs. energy efficiency.

4. Empirical Evaluation and Performance Benchmarks

Comprehensive experiments consistently demonstrate the practical superiority of FedPNS methods in diverse operational regimes:

Scenario	Baselines	Key FedPNS Gains
Wireless FL, non-IID (Marnissi et al., 2024)	Uniform, greedy, rounding	3–5× speedup, 4–6× energy reduction, +10% accuracy
Gradient-guided FL (Wu et al., 2021)	FedAvg, top-$p_i$9 selection	30–50% fewer rounds, +1–3% final accuracy
GPU/satellite SLO (Stanisic et al., 11 Nov 2025)	Hard-deadline heuristics	+13.77% SLO compliance, 72.5% less waste
Asynchronous FL (Yang et al., 2023)	Random, greedy	Higher accuracy for fixed energy, improved fairness

Experiments on MNIST and CIFAR-10 with hundreds to thousands of clients consistently show that probabilistically optimized participation schedules lead to faster accuracy attainment, dramatically reduced energy use, and fairer distribution of communication workload. Deterministically excluding rare or expensive clients frequently leads to plateaued or reduced accuracy (Marnissi et al., 2024).

5. Extensions and Specializations

FedPNS concepts have been extended to more demanding distributed and privacy-conscious regimes:

• Decentralized/Personalized Learning: Nodes employ cryptographically secured aggregation (e.g., Shamir secret sharing, SecAgg) in decentralized settings, and select subsets of peers using correlated adversarial multi-armed bandits with Tsallis-Inf weights. Correlation structure among arms (client subsets) is leveraged for regret minimization, yielding tight $i$0 regret bounds (Zec et al., 2023).
• Asynchronous and Straggler-Prone Networks: Clients operate asynchronously, submitting updates probabilistically per round, with minimum-selection guarantees to mitigate starvation. Bandwidth allocations are jointly assigned, and optimization utilizes fractional programming (Yang et al., 2023).
• GPU-Accelerated / 3D Continuum: Client selection is based on analytical latency models, treating efficiency as a random variable, thus enabling operation without historical monitoring and in environments with rapidly varying hardware/network conditions (e.g., satellites, mobile, edge-cloud) (Stanisic et al., 11 Nov 2025).
• Heterogeneous Fairness Metrics: Objective weights $i$1 or minimum required participation probabilities can encode fairness or regulatory criteria.

6. Comparative Analysis and Limitations

FedPNS is consistently superior to baseline heuristics, deterministic selection, and random scheduling by all principal performance metrics (energy, time, final accuracy, resource fairness). However, certain caveats and open questions persist:

• Local Optima: While convergence to stationary points is guaranteed, global optimality is not always assured due to the non-convex coupling between continuous probability and resource/power variables.
• Dynamics and Feedback: Highly dynamic environments, especially with time-variant channels or workloads, may require real-time model updates or robust estimators for certain parameters (e.g., SLO compliance in absence of historical $i$2) (Stanisic et al., 11 Nov 2025).
• Privacy and Decentralization: Implementation in privacy-critical or weakly connected topologies necessitates secure aggregation (SecAgg) and decentralized coordination protocols (Zec et al., 2023).
• Non-IID and Data Skew: The full benefit of gradient-alignment-guided FedPNS is realized under pronounced non-IID data heterogeneity; in near-IID regimes, simpler schemes may suffice, though FedPNS remains competitive.

7. Prospects and Applications

FedPNS is a foundational tool for scalable, robust, and resource-efficient distributed learning in emerging applications:

• Massive Device Federated Learning: Wireless cellular, IoT, and industrial settings with energy and latency heterogeneity.
• Zero/Historical-Free Scheduling: Environments lacking reliable device state monitoring (e.g., space networks).
• Privacy-Preserving Personalized Learning: Decentralized edge learning with privacy constraints and adaptive peer discovery.
• Multi-cell and Interference-Limited Scenarios: Extension to settings where SINR constraints and inter-cell coordination are required.

Future directions include tighter integration with joint uplink-downlink design, adaptive fairness criteria, and expanded applicability in highly non-stationary or adversarial environments (Marnissi et al., 2024, Yang et al., 2023).