Training-Free Adaptive Bandwidth Allocation

• Training-free adaptive bandwidth allocation is a method that dynamically assigns bandwidth using instantaneous resource counters and fixed per-class thresholds without relying on historical training.
• It leverages multi-level call admission control, priority-based schemes, and rule-driven BAM frameworks to maintain high utilization, reduce blocking, and ensure negligible handover drop rates.
• These approaches are applied across wireless networks and compute systems, demonstrating robust performance and rapid responsiveness to dynamic traffic loads and application requirements.

Training-free adaptive bandwidth allocation encompasses a class of resource management algorithms for networking and compute systems designed to maximize performance metrics such as bandwidth utilization, blocking and dropping probabilities, and service differentiation. These methods allocate bandwidth resources dynamically based on instantaneous system state, priority, and application requirements, but fundamentally avoid the use of offline traffic training, historical modeling, or learned predictors. Decision rules depend only on current occupancy, resource counters, and application-level constraints.

1. Foundational Principles and Key Models

The core principle of training-free adaptive bandwidth allocation is on-the-fly adaptation tied to per-class thresholds and resource counters, without the need to estimate traffic distributions or train statistical predictors. Pioneering implementations include multi-level bandwidth-adaptive Call Admission Control (CAC) for wireless networks (Chowdhury et al., 2014, Chowdhury et al., 2015), priority-based multi-class adaptation for differentiated services (Chowdhury et al., 2014), bidirectional sharing protocols in MPLS DS-TE bandwidth allocation models (e.g., AllocTC-Sharing (Reale et al., 2019)), rule-driven autonomic frameworks via generalized BAMs (GBAM) (Reale et al., 2018), stochastic allocation models for D2D wireless networks (Baccelli et al., 2021), iterative optimization in multi-operator spectrum sharing (George et al., 10 Jun 2025), and self-supervised gating for adaptive compute allocation in transformer models (Sim, 31 Dec 2025).

A defining feature is the use of instantaneous per-class counters (e.g., number of current calls $N_m$, bandwidth-in-use $B_{m,a}$) and rigid per-class adaptation thresholds (e.g., minimum allocation $B_{m,n}$ for new calls, $B_{m,h}$ for handovers), allowing protocol decisions without historical context or learning phase.

2. Multi-Level Adaptive Allocation in Wireless Networks

Training-free multi-level adaptive CAC algorithms are specified for wireless networks that support heterogeneous traffic types (real-time vs. non-real-time) (Chowdhury et al., 2014, Chowdhury et al., 2015). The admission control logic is outlined as follows:

• On each call arrival, compute currently occupied bandwidth $C_\text{occ} = \sum_{m=1}^M N_m B_{m,a}$.
• If sufficient free bandwidth exists, admit at full requested rate $B_{m,r}$.
• Otherwise, for adaptive (non-real-time) classes, compute total releasable bandwidth via degradation:
  • $$C_\text{rel,new} = \sum_{k=q+1}^M N_k (B_{k,a} - B_{k,n})$$
  • $$C_\text{rel,hand} = \sum_{k=q+1}^M N_k (B_{k,a} - B_{k,h})$$
• Admit new/handover calls by degrading existing adaptive calls just enough to reach per-class minimum allocations $B_{k,n}$, $B_{k,h}$.
• If these tests fail, block (new) or drop (handover) the call.

The parameters $B_{m,n}$ and $B_{m,h}$ are set offline and enforce service differentiation: handover calls receive stricter protection (more bandwidth can be degraded) than new calls, driving handover dropping probabilities negligible while maintaining high utilization. All choices are made instantaneously and require no traffic model estimation.

Markov chain and birth-death queueing models formalize state dynamics, with analytical steady-state probabilities enabling precise blocking ($P_B$) and dropping ($P_D$) calculations:

$$P(i) = \begin{cases} \frac{(\lambda_n + \lambda_h)^i}{i! \mu^i} P(0) & 0 \leq i \leq N \ \frac{(\lambda_n + \lambda_h)^N \lambda_h^{i-N}}{i! \mu^i} P(0) & N < i \leq N + S \end{cases}$$

$P_B = \sum_{i=N+L}^{N+S} P(i), \qquad P_D = P(N+S)$

Empirical results confirm that even under heavy load ($\lambda_n > 0.25$ calls/s), handover dropping rates $P_D$ are below $5 \times 10^{-4}$ and bandwidth utilization remains above 0.94 (Chowdhury et al., 2014, Chowdhury et al., 2015).

3. Priority-Based and Service-Differentiated Schemes

Training-free priority-based schemes allocate bandwidth adaptively with strict ordering by traffic class, call type, or application priority (Chowdhury et al., 2014). Bandwidth degradation factors $y_{m,p}$ are specified for each priority $p$ and traffic class $m$, defining maximum fractional release per call.

Admission occurs by:

• Releasing bandwidth from higher-priority classes first.
• Solving $F + \sum_{m=1}^M R_m(p) \geq C_{p,r}$ for the requested call, with $R_m(p)$ the total releasable from class $m$ for priority $p$.

This approach reduces blocking/dropping for prioritized traffic and preserves overall utilization. Analysis shows utilization $U$ remains high and blocking probability $P_B(p)$ strictly decreases with increasing priority. No statistical learning or historical data are required; all decisions are made from the instantaneous state vector $\{i_m, C_{m,a}\}$ (Chowdhury et al., 2014).

4. Training-Free Bandwidth Allocation Models in MPLS/DS-TE Networks

Bandwidth Allocation Models (BAMs) traditionally use rigid class-level reservations (e.g., MAM) or hierarchical sharing (e.g., RDM, G-RDM). AllocTC-Sharing (Reale et al., 2019) and generalized BAMs (GBAM) (Reale et al., 2018) extend this framework by enabling bidirectional sharing (high-to-low and low-to-high loans).

Key allocation formula for AllocTC-Sharing:

$$G_i(t) = N_i(t) + \text{free}_i(t) + \sum_{h>i} \mathrm{loan}^LTH_{h \to i}(t) + \sum_{l<i} \mathrm{loan}^{HTL}_{l \to i}(t)$$

Where $BC_i$ is the reserved bandwidth for class $i$, $N_i(t)$ instantaneous usage, and "loan" terms denote available slack from other classes. Borrowed resources remain preemptible for the donor class. No traffic prediction or learning phase is necessary; adaptation is continuous and immediate.

The GBAM framework further enables rule-based autonomy. By monitoring live metrics (utilization $U$, preemption $P$, blocking $Bk$), it switches configuration via high-level policies in a MAPE (Monitor–Analyze–Plan–Execute) loop. For instance:

• If $U_\text{obs} < U_\text{min}$, switch to aggressive sharing.
• If $P_\text{obs} > P_\text{max}$, switch to conservative mode.

Such rule-driven adaptive switching achieves intermediate operating points and does not require offline training (Reale et al., 2018).

Simulation results indicate that adaptive switching yields a mean utilization $U \approx 85\%$ and preemptions $\bar{P} \approx 43$/hour—substantially improving over static models. All adaptation leverages only live system metrics (Reale et al., 2018, Reale et al., 2019).

5. Stochastic, Probabilistic, and Semi-Static Adaptive Techniques

A class of training-free allocation models utilize stochastic geometry or optimization to maximize metrics such as average throughput, fairness, or percentiles of QoE, without data-driven modeling.

In D2D wireless networks (Baccelli et al., 2021), transmitters independently select user type (bandwidth demand) and randomly choose subbands ("chunks") per allocation rules. The performance metrics—success probability, meta-distribution of SIR, and Shannon throughput—are derived analytically:

$$P_s^{(k)}(\theta) = \prod_{i=1}^K \exp(-\lambda p_i C \theta^\delta \sum_{t} p_{k,i}^{(t)} (t/k)^\delta)$$

A major insight is that, given equal mean signal and interference power, networks with higher traffic variability achieve strictly better performance metrics; no learning or historical estimation is necessary.

For terahertz multi-connectivity systems, adaptive sub-band bandwidth allocation uses offline curve-fitting and convex optimization (SCA) (Shafie et al., 2021). The entire process operates via successive convex programs, using real-time topology and offline channel parameters, obviating pilot training or online learning. Throughput gains of 13–33% over equal-bandwidth designs are obtained, and all adaptation is computational (not learning-based).

Semi-static bandwidth sharing for real-time video traffic operates in hyperperiods; resource allocation and spectrum sharing are updated by solving convex subproblems and leveraging Lyapunov drift-plus-penalty control (George et al., 10 Jun 2025). The method provably achieves $O(1/V)$ optimality gap, responds rapidly to traffic and channel changes, and requires only infrequent coordination. Absence of historical model fitting sharply distinguishes the approach from ML/RL-based methods.

6. Adaptive Compute Allocation: Training-Free Gating

Beyond networking, training-free adaptive allocation principles have been applied to compute resources in AI inference. PonderTTT (Sim, 31 Dec 2025) is a fully training-free gating strategy for selectively triggering Test-Time Training (TTT) updates in transformer models. The update schedule is computed via self-supervised reconstruction loss and a calibrated threshold $\tau$:

• Initial $\tau$ set as the $(1-\rho)$-quantile of probe losses
• Online adjustment via an EMA loop; no auxiliary classifier

This yields 82–89% Oracle Recovery at half the dense TTT cost (FLOPs), and up to 16% lower loss on OOD tasks, without any learned signal, traffic traces, or historical training.

7. Computational Workflow and Practical Considerations

Across all models, training-free adaptive bandwidth allocation exhibits consistent workflow properties:

• All decisions depend directly on instantaneous resource occupancy and a fixed configuration of per-class thresholds or priority scalars.
• No offline traffic or traffic distribution estimation is ever performed or required.
• Computational overhead is minimal, dominated by simple arithmetic, priority sorting, or per-frame convex optimization, enabling microsecond admission decisions in real-time settings.
• Adaptation to traffic spikes, changing loads, or application requirements is immediate, governed solely by current counters and rules.
• Offline parameter setting (e.g., thresholds $B_{m,n}$, $B_{m,h}$, degradation factors $y_{m,p}$) generally occurs once per class or per system deployment.

Empirical and simulation results consistently demonstrate that such methods can drive critical metrics (handover dropping, blocking rates, utilization) into near-optimal regimes (e.g., $P_D < 5\times10^{-4}$, $U > 0.95$) (Chowdhury et al., 2014, Chowdhury et al., 2015, Chowdhury et al., 2014).

8. Comparative Merits and Generalization

Training-free adaptive bandwidth allocation is robust to input traffic variation, system disturbance, and novel service mixes, owing to its lack of reliance on historical learning or predictive estimation. The class of methods encompasses multi-level CAC, differentiated adaptation, bidirectional sharing BAMs, rule-based autonomic frameworks, stochastic geometry, convex optimization, and inference-time gating, unifying them under the shared principle of instantaneous, configuration-driven control.

A plausible implication is that future deployment in highly dynamic or heterogeneous environments will favor training-free adaptive allocation for its analytical tractability, rapid responsiveness, and zero dependency on site-specific model fitting.

---

References

• Multi-level bandwidth-adaptive CAC for wireless networks (Chowdhury et al., 2014, Chowdhury et al., 2015)
• Priority-based adaptation for multi-class wireless traffic (Chowdhury et al., 2014)
• AllocTC-Sharing MPLS BAM (Reale et al., 2019)
• Rule-based autonomic allocation with GBAM (Reale et al., 2018)
• Stochastic geometry models for service differentiation (Baccelli et al., 2021)
• Iterative semi-static sharing for QoE in wireless operators (George et al., 10 Jun 2025)
• Adaptive compute allocation via self-supervised gating (Sim, 31 Dec 2025)
• Training-free sub-band allocation in THz networks (Shafie et al., 2021)