Human–Machine Coordinated DBA

• HMC‑DBA is a dynamic framework where human-defined policies guide automated controllers in real-time bandwidth allocation.
• It leverages diverse substrates like MPLS/DS-TE, SDN/OpenFlow, and P4 hardware to execute precise actions such as admission, preemption, and prediction.
• The system balances efficiency and stability by managing trade-offs through feedback loops, predictive analytics, and controlled reconfiguration.

Searching arXiv for the cited HMC-DBA and related dynamic bandwidth allocation papers to ground the article in current literature. Human–Machine Coordinated Dynamic Bandwidth Allocation (HMC‑DBA) denotes a class of bandwidth-control schemes in which humans, network managers, or higher-level policy systems define objectives, tolerances, service classes, and operating constraints, while machine-side controllers execute fine-grained allocation, admission, prediction, rate assignment, preemption, or reconfiguration in real time. The expression appears explicitly in work on predictive XR transport for human-to-machine collaboration, where head-movement prediction drives both camera pre-orientation and proactive bandwidth grants (Mondal et al., 21 Jul 2025). Closely related literature in MPLS/DS‑TE, SDN/OpenFlow, xDSL/femtocell access, PON DBA, and wireless resource allocation follows the same division of labor: operators define SLAs, thresholds, and sharing rules, and automated components implement detailed control actions at packet, flow, LSP, or cycle timescales (Torres et al., 2021, Chowdhury et al., 2015, Oliveira et al., 2019).

1. Conceptual basis and scope

In the literature, HMC‑DBA is grounded in a recurring architectural split between policy formation and operational control. On the human side, managers specify service differentiation, risk tolerance, and acceptable trade-offs among utilization, blocking, preemption, devolution, latency, jitter, or secrecy. On the machine side, controllers execute per-request or per-cycle decisions such as admission, denial, borrowing, preemption, throttling, and queue or meter configuration. This separation is explicit in Bandwidth Allocation Model (BAM) work, where the human side selects bandwidth constraints and sharing rules, and the machine side enforces admission control, loans, and preemption dynamically (Reale et al., 2019).

The same pattern appears in SLA-oriented access control. In femtocell backhaul, a Bandwidth Broker maintains SLA policy, monitors current and historical traffic, computes a dynamic reservation, and enforces rate control, while operators determine the SLA, the observation window, and throttling rules (Chowdhury et al., 2015). In cognitive BAM management, the network manager defines tolerances for preemption, devolution, and blocking, while a Case-Based Reasoning (CBR) engine decides when to switch among MAM, RDM, and ATCS and learns from positive and negative outcomes (Oliveira et al., 2019). This suggests that HMC‑DBA is less a single algorithm than a coordination pattern that links operator intent to automated bandwidth control.

A second defining trait is heterogeneity of control substrate. HMC‑DBA has been implemented or approximated over MPLS/DS‑TE with BAMs, SDN/OpenFlow flow rules, Bandwidth Broker rate control, PON DBA cycles, P4-programmable hardware, and learned wireless allocators (Torres et al., 2021, Mafioletti et al., 2023, Hao et al., 2023). The common abstraction is not the network technology, but the closed-loop relation between policy, observation, and adaptation.

2. Control loop and architectural organization

Representative systems implement HMC‑DBA as a monitored feedback loop. BAMSDN decomposes the controller into Admission, BAM, State of the Network, and Execution modules. New flows arrive through OFPT_PACKET_IN, are classified into traffic classes, checked against per-link and per-class constraints, and are then granted, denied, or granted with preemption; the resulting action is translated into FLOW_MOD ADD, FLOW_MOD DELETE, or drop behavior (Torres et al., 2021). The femtocell Bandwidth Broker implements a similar loop of monitoring, database update, reservation computation, rate-control configuration, and re-monitoring (Chowdhury et al., 2015). BAMCBR organizes the cycle as retrieve, reuse/adapt, revise, and retain, with separate positive and negative case databases (Oliveira et al., 2019).

These systems expose different but structurally similar coordination surfaces.

System	Human-side levers	Machine-side actions
BAMSDN	BAM mode, CT definitions, BCs, hard/soft reconfiguration	classify flows, check per-link constraints, grant/deny/preempt, install OpenFlow rules
Bandwidth Broker	SLA, $T$, $t_1$, throttling limits	compute $B_R(t)$, derive $B_B(t)$, enforce rate control
BAMCBR	tolerances, policy tuples, optional proposed solution	retrieve/adapt/revise/retain cases, switch BAM behavior
Split vDBA/P4	low-latency Alloc-ID policy, reserved capacity	intercept DBRu, buffer requests, modify BWMAP in hardware

A related architectural development is the split DBA design for virtual PONs. A software virtual DBA computes a baseline BWMAP while intentionally leaving some capacity unallocated; a P4-programmable SmartNIC then intercepts low-latency DBRu reports, buffers them locally, intercepts the downstream BWMAP, and inserts grants for low-latency traffic into the unused region at line rate (Mafioletti et al., 2023). Although no human interface is formalized there, the mapping of T‑CONTs or Alloc‑IDs to low-latency treatment and the amount of reserved capacity are natural policy hooks.

3. Allocation models and control actions

A central class of HMC‑DBA mechanisms is BAM-based class sharing in DS‑TE. In MAM, each traffic class has private bandwidth only. For a link $e$, the constraints are class-local, for example

$$\sum_{l \in CT_0,\; l \ni e} bw_l \leq BC_{e,0}, \quad \sum_{l \in CT_1,\; l \ni e} bw_l \leq BC_{e,1},$$

with no borrowing (Torres et al., 2021). In RDM, bandwidth constraints are nested, such as

$$\sum_{l \in (CT_0 \cup CT_1 \cup CT_2),\; l \ni e} bw_l \leq BC_{e,0}, \quad \sum_{l \in (CT_1 \cup CT_2),\; l \ni e} bw_l \leq BC_{e,1},$$

allowing hierarchical sharing with preemption (Torres et al., 2021). AllocTC‑Sharing extends this by permitting both low-to-high and high-to-low borrowing through temporary loans while still preserving bandwidth constraints for traffic classes (Reale et al., 2019).

The operational actions induced by these models are grant, deny, grant with preemption, and reconfiguration. BAMSDN also distinguishes hard and soft bandwidth-constraint reconfiguration. In hard mode, a new $BC'_{e,i}$ is applied immediately and the controller finds a set $\mathcal{P}$ of LSPs to remove so that the new constraints hold. In soft mode, the new bandwidth constraints become admission limits for future requests but do not force immediate teardown of current allocations (Torres et al., 2021). The trade-off is direct: hard reconfiguration gives fast policy effect but can cause abrupt preemptions; soft reconfiguration preserves ongoing flows but delays convergence.

Another control family is history-based reservation. In the femtocell Bandwidth Broker model, the reserved bandwidth is the moving average of recent femtocell demand: $$B_R(t) = \frac{1}{N}\sum_{n=m}^{m+N} B_F(t - n t_1),$$ and, when current availability $t_1$0 is insufficient, the broker computes borrowing bandwidth

$t_1$1

and throttles non-femtocell traffic accordingly (Chowdhury et al., 2015). This is an explicit instance of policy-bounded automation: the human side sets $t_1$2, $t_1$3, and permissible limits, while the broker controls exact reservation and borrowing.

Tier-based allocation provides a third pattern. In semi-synchronous federated learning, clients are assigned to latency-based tiers $t_1$4, bandwidth is allocated across tiers under $t_1$5, and the effective objective is to maximize weighted processed samples with higher weights for lower-latency tiers (Yu et al., 2024). This suggests an HMC‑DBA generalization in which tier or slice assignment acts as a coarse human-readable policy layer, while workload and bandwidth within tiers are machine-optimized.

4. Prediction, cognition, and learned machine-side control

The most explicit HMC‑DBA formulation couples bandwidth control to human state prediction. In immersive H2M XR collaboration, head orientation is modeled through quaternions, average head rotation speed is computed as

$t_1$6

and BiLSTM predictors estimate future head motion over horizons up to $t_1$7 ms (Mondal et al., 21 Jul 2025). The predicted angular displacement then drives a model of effective XR bitrate,

$t_1$8

so that upcoming XR frame size and bandwidth demand can be predicted before the frame is generated (Mondal et al., 21 Jul 2025). The resulting HMC‑DBA scheme allocates grants based on $t_1$9, $B_R(t)$0, and $B_R(t)$1, rather than only on reactive queue reports.

Learning-based bandwidth control also appears in more generic settings. A graph-neural-network policy can represent per-user allocations with a parameter count that does not change with the number of users, and hybrid-task meta-learning can meta-train that policy across varying QoS requirements, channels, and resource budgets. In the reported results, the HML approach improves the initial performance by $B_R(t)$2 and sampling efficiency by $B_R(t)$3 relative to the compared benchmarks, while approaching the optimal iterative policy after fine-tuning (Hao et al., 2023). This provides a machine-side allocator that is structurally compatible with operator-defined slice budgets, QoS thresholds, and priority mappings.

Cognitive reconfiguration through experience is another machine-side pattern. BAMCBR retrieves similar cases, adapts or proposes a BAM switch, tests the outcome, and stores the result as a positive or negative case (Oliveira et al., 2019). The paper’s reported totals show that BAMCBR can reduce preemption to $B_R(t)$4 and devolution to $B_R(t)$5 while keeping $B_R(t)$6 unbroken LSPs, compared with $B_R(t)$7 preemptions under static RDM and $B_R(t)$8 devolutions under static ATCS (Oliveira et al., 2019). In this formulation, human coordination enters through tolerances, policies, and optional solution proposals.

Finally, semantic compression can be interpreted as a machine-side DBA primitive. In hybrid event-RGB transmission, the system disentangles image-specific, event-specific, and shared latents and minimizes a multi-task rate–distortion objective of the form

$B_R(t)$9

so that more symbols are allocated to events for dynamic details or to images for static information (Yang et al., 25 Jun 2025). The paper does not introduce humans explicitly, but it provides a clear machine-side rate allocator whose weights are plausible policy knobs.

5. Implementation substrates and timing-sensitive realization

HMC‑DBA has been realized over several distinct execution substrates. In MPLS-over-SDN, BAMSDN uses POX $B_B(t)$0, OpenFlow $B_B(t)$1, Mininet, and iPerf3, while preserving an MPLS/DS‑TE abstraction externally and implementing LSPs internally as Software Switched Paths (Torres et al., 2021). Admission and preemption decisions are enforced through flow-table operations and port-level queues or meters, making runtime class-based bandwidth control programmable at the controller.

Timing-sensitive access networks introduce another dimension: scheduler correctness. In EPON, inaccurate RTT estimates distort the relation between the end of ONU $B_B(t)$2’s burst and the start of ONU $B_B(t)$3’s burst, with

$B_B(t)$4

If $B_B(t)$5, collisions occur; if $B_B(t)$6, bandwidth is wasted (Hong et al., 2014). Under RTT deviation $B_B(t)$7, the paper reports an average collision rate of about $B_B(t)$8, average waste of transmission time around $B_B(t)$9, and average utilization about $e$0; adding a complement $e$1 reduces collision rate to $e$2 and yields average line utilization of $e$3 (Hong et al., 2014). This is not framed as human coordination, but it exposes a direct control knob between robustness and efficiency.

A complementary EPON design, Hybrid Slot-Size/Rate, fixes one high-priority slot per ONU in each $e$4 ms frame while assigning the dynamic part to best-effort traffic. The architecture reduces high-priority packet delay to about $e$5, compared with about $e$6 under conventional slot-size DBA, and keeps the standard deviation of high-priority packet delay below $e$7 (An et al., 2014). The fixed HP allocation versus dynamic BE allocation is a clear example of separating stable policy reservations from opportunistic machine-side filling.

Programmable hardware further compresses the control loop. In virtual PONs, P4 hardware executes low-latency DBA operations at microsecond scale by intercepting DBRu and editing BWMAPs on the fly, while slower virtual DBA logic in software maintains the broader multi-service or multi-tenant policy (Mafioletti et al., 2023). In cloud stream analytics, SDN-based controllers use application-layer queue information, sender and receiver backlogs, and per-link capacities to compute cross-layer flow rates at runtime and improve both throughput and average tuple latency relative to TCP’s flow-rate fairness (Aljoby et al., 2018). Across these domains, HMC‑DBA repeatedly appears as a hierarchy: slow policy, fast enforcement.

6. Trade-offs, limitations, and open problems

The dominant HMC‑DBA trade-off is between efficiency and disruption. In BAM literature, more sharing improves utilization and often reduces blocking, but also increases the probability of preemptions or devolutions. MAM gives isolation and no sharing; RDM yields high utilization with preemption; AllocTC‑Sharing pushes utilization further but allows both HTL and LTH loans, with correspondingly higher operational volatility (Reale et al., 2019). Autonomic switching with G‑BAM makes this trade-off explicit: switching among BAM behaviors based on high-level management rules can improve link utilization and preemption relative to static choices, but only by moving along a nontrivial utilization–stability frontier (Reale et al., 2018).

Prediction accuracy is a second critical limit. In XR HMC‑DBA, under-prediction of frame size leads to additional queueing cycles and jitter, while over-prediction wastes bandwidth; the latency model therefore depends directly on the prediction residual $e$8 (Mondal et al., 21 Jul 2025). In wireless meta-learned allocation, transferability depends on the relation between meta-training and meta-testing task distributions; the paper reports strong adaptation, but still distinguishes initial performance from fine-tuned performance under unseen scenarios (Hao et al., 2023). In cognitive BAM management, early stages without prior cases can try poor solutions before the system learns stable responses (Oliveira et al., 2019).

A third issue is complexity. Some near-optimal DBA mechanisms are computationally expensive, while more scalable schemes are approximate. In bandwidth-limited sensor tracking, exhaustive search is combinatorial, convex relaxation is cubic in the problem size, and approximate dynamic programming is much cheaper while remaining close in MSE performance to exhaustive search and outperforming greedy and nearest-neighbor allocation (Masazade et al., 2011). This suggests that HMC‑DBA systems must often choose not just an allocation, but an allocation algorithm whose latency and explainability match the operational context.

Several papers also note structural limitations: centralized controllers may face scaling and monitoring overhead; some systems assume perfect or timely telemetry; some policy parameters are static at runtime; and some formulations optimize only a single bottleneck or a single resource type (Chowdhury et al., 2015, Aljoby et al., 2018, Yu et al., 2024). A plausible implication is that future HMC‑DBA research will continue to move toward hierarchical control, richer policy languages, explicit human override paths, and multi-objective formulations that combine utilization, fairness, latency, jitter, preemption tolerance, and interpretability rather than optimizing any one of them in isolation.