DQRoute: Adaptive Routing & Recognition

• DQRoute is an overloaded term defining deep Q-learning routing strategies in quantum/classical networks and a long-tailed recognition framework that fuses expert outputs.
• In routing applications, methods like QuDQN use difficulty-aware masked policies and centralized state snapshots to resolve network requests efficiently.
• In visual recognition, DQRoute employs a divide, weight, and route paradigm that dynamically fuses general, medium, and tail experts to boost rare class performance.

DQRoute denotes two distinct usages in the recent arXiv literature. In one usage, it is the exact name of “Divide, Weight, and Route,” a long-tailed recognition framework that combines difficulty-aware optimization with dynamic expert fusion (Wei et al., 27 Aug 2025). In routing research, the same label is used more loosely for deep Q-learning-based routing systems; within that sense, the closest direct technical match is QuDQN, a centralized adaptive entanglement-routing-and-resource-allocation method for repeater-based quantum networks that the authors explicitly distinguish from the exact name “DQRoute” while placing it in the same general problem space as DQN-driven quantum routing (Jallow et al., 1 Mar 2025).

1. Terminological scope and disambiguation

The naming distinction is fundamental. The paper “Adaptive Entanglement Routing with Deep Q-Networks in Quantum Networks” proposes QuDQN, not “DQRoute,” and states that the strict answer to whether the method is literally DQRoute is no; at the same time, it also states that QuDQN is “exactly the same family of approach one would mean by a DQN-based adaptive quantum routing scheme,” and explicitly cites prior work named DQRA: Deep quantum routing agent for entanglement routing in quantum networks (Jallow et al., 1 Mar 2025). By contrast, “Divide, Weight, and Route” uses DQRoute as its formal method name in long-tailed visual recognition rather than in communication routing (Wei et al., 27 Aug 2025).

A useful editorial consolidation is “DQRoute-style routing” (Editor's term): a family of deep Q-learning routing methods in which a learned Q-function scores feasible routing actions from network state. In the provided literature, this family includes centralized entanglement-routing systems for quantum repeater networks, distributed or hierarchical packet-routing systems, and heterogeneous wireless routing methods that extend the action space to communication technology and subband choices (Jallow et al., 1 Mar 2025).

This naming overlap creates a recurring misconception. In quantum-network discussions, “DQRoute” often refers conceptually to a DQN-based adaptive routing mechanism even when the actual paper-specific name is different. In vision, however, DQRoute is not a router over network paths at all, but a modular mixture-of-experts recognition framework. Any technical use of the term therefore requires domain qualification.

2. Reinforcement-learning formulation in quantum-network DQRoute-style systems

QuDQN formulates an adaptive entanglement routing and resource allocation problem on a repeater-based quantum network modeled as an undirected graph $G=(V,E)$. At each time slot $t$, the agent observes the current topology $G^{(t)}=(V,E^{(t)})$, node qubit capacities $C^{(t)}$, pending communication demands $D^{(t)}$, edge entanglement-generation success probabilities $p_e$, node swapping success probabilities $q_v$, and a fidelity requirement $F$. A request is treated as a source-destination entanglement demand whose feasibility depends on path feasibility, qubit resources, success probabilities, and fidelity threshold (Jallow et al., 1 Mar 2025).

The action semantics are higher-level than elementary quantum control. At each time slot, the agent “selects a request from the set of pending requests $D^{(t)}$ and assigns a path connecting the source and destination nodes.” This makes the method a single-request-at-a-time scheduling plus path selection framework over a batch of pending requests, rather than a multipath flow optimizer or an explicit entanglement-swapping scheduler. The state abstraction is a centralized global snapshot of network resources and request backlog; the paper specifies the semantics of the input state, but not a tensorized encoding, feature-vector dimension, adjacency format, or graph embedding (Jallow et al., 1 Mar 2025).

The action space is described as the set of feasible routing schedule decisions $a \in \mathcal{A}$. The paper further introduces a masked action policy excluding invalid actions, such as routing through paths with insufficient qubits or failed swaps. That mask is one of the most reusable technical ideas in DQRoute-style formulations because it enforces feasibility structurally rather than through pure penalty shaping. The printed action-selection equation is malformed, and a later masked-policy equation is closer to ordinary DQN action selection, so the paper is clearer at the semantic than the symbolic level (Jallow et al., 1 Mar 2025).

The reward is the most explicit component of the formulation:

$t$0

where $t$1 is the number of resolved requests after step $t$2, $t$3 is the total number of requests in the episode, $t$4 rewards resolved requests, $t$5 penalizes incomplete episodes, $t$6 indicates whether $t$7 is the final step, $t$8 is a discount factor, and $t$9 weights the fidelity/success term. The paper itself notes a notational inconsistency: the prose mentions $G^{(t)}=(V,E^{(t)})$0 while the printed equation contains $G^{(t)}=(V,E^{(t)})$1. The high-level design is therefore clear, but the printed formula is imperfect (Jallow et al., 1 Mar 2025).

Training uses standard DQN machinery with Bellman targets, temporal-difference error, and a target network:

$G^{(t)}=(V,E^{(t)})$2

$G^{(t)}=(V,E^{(t)})$3

$G^{(t)}=(V,E^{(t)})$4

The target network is updated every $G^{(t)}=(V,E^{(t)})$5 steps,

$G^{(t)}=(V,E^{(t)})$6

which identifies the method as a standard DQN-style approach with target-network stabilization rather than an explicitly described Double DQN or dueling architecture. The paper does not specify replay-buffer details, exploration schedule, optimizer, or exact neural architecture, so QuDQN is technically informative but not fully reproducible from the text alone (Jallow et al., 1 Mar 2025).

3. Quantum-network model, implementation choices, and reported behavior

QuDQN assumes a repeater-based quantum network over an undirected graph. Nodes are quantum repeaters or devices, edges are lossy optical links capable of elementary entanglement generation, and connected nodes are assumed to be within an entanglement threshold of approximately $G^{(t)}=(V,E^{(t)})$7 km. Each node has finite qubit capacity $G^{(t)}=(V,E^{(t)})$8; in experiments, node qubit counts are fixed at $G^{(t)}=(V,E^{(t)})$9 in some grid-scaling scenarios and $C^{(t)}$0 in larger benchmark comparisons. Channel capacities are distributed between $C^{(t)}$1 and $C^{(t)}$2, entanglement generation succeeds with probability $C^{(t)}$3, swapping succeeds with probability $C^{(t)}$4, link Bell-pair fidelity is sampled as $C^{(t)}$5, and the minimum end-to-end fidelity requirement is $C^{(t)}$6. Decoherence, memory lifetime, and temporal decay are not explicitly modeled beyond general motivation, and purification is not part of QuDQN itself (Jallow et al., 1 Mar 2025).

The transition logic includes explicit qubit-consumption semantics: intermediate nodes deduct $C^{(t)}$7 qubits only if entanglement swapping succeeds, while end nodes deduct $C^{(t)}$8 qubit. This means the environment update depends on path feasibility and stochastic link and swapping success probabilities, but no explicit transition kernel $C^{(t)}$9 is written. The model is centralized, because the agent observes a global graph, all node capacities, and the full demand set, and outputs a global routing schedule (Jallow et al., 1 Mar 2025).

The implementation stack is conventional: Python, NetworkX for graph modeling and shortest paths, NumPy for numerical work, and TensorFlow for the DQN. Hyperparameters explicitly reported include learning rate $D^{(t)}$0, mini-batch size $D^{(t)}$1, reward coefficients $D^{(t)}$2, $D^{(t)}$3, $D^{(t)}$4, $D^{(t)}$5, and target-network update interval $D^{(t)}$6. The network is described only semantically as taking “network topology, qubit capacity, requests” as input and producing “routing schedules, path assignments” as output; no hidden-layer widths, graph layers, or activations are specified (Jallow et al., 1 Mar 2025).

Evaluation uses two baseline families. Internal variants replace learned routing with shortest-path or random request-order routing: QuDQN-Shortest and QuDQN-Random. External baselines are EFiRAP and ERDREG-PU. Reported metrics are request completion, throughput, quantum channel utilization, and qubit utilization. In $D^{(t)}$7 to $D^{(t)}$8 grid topologies with $D^{(t)}$9 qubits per node and $p_e$0–$p_e$1 requests, QuDQN resolves about $p_e$2 of requests, versus roughly $p_e$3 for QuDQN-Random and $p_e$4 for QuDQN-Shortest. In the $p_e$5 grid it exceeds QuDQN-Shortest by $p_e$6 and QuDQN-Random by $p_e$7; in the $p_e$8 grid, the advantage is $p_e$9 and $q_v$0, respectively. In terms of qubit usage, QuDQN uses $q_v$1 fewer qubits than QuDQN-Random and $q_v$2 fewer than QuDQN-Shortest overall (Jallow et al., 1 Mar 2025).

Against EFiRAP and ERDREG-PU, in a $q_v$3 grid with $q_v$4 qubits per node and $q_v$5–$q_v$6 source-destination pairs, QuDQN is reported to resolve $q_v$7 of requests at all demand levels. Throughput improvements are $q_v$8 over EFiRAP and $q_v$9 over ERDREG-PU at $F$0 requests, widening to $F$1 and $F$2 at $F$3 requests. At $F$4 requests, node qubit utilization is $F$5 for QuDQN versus $F$6 for EFiRAP and $F$7 for ERDREG-PU; at $F$8 requests, it is $F$9 versus $D^{(t)}$0 and $D^{(t)}$1. These are simulation results under a simplified physical model, and the paper explicitly lacks ablations that would isolate the effect of the fidelity term, masking, or target-update interval (Jallow et al., 1 Mar 2025).

4. DQRoute-style routing across classical packet and wireless networks

The DQRoute-style family extends well beyond entanglement routing. “Hierarchical Deep Double Q-Routing” proposes a hierarchical cluster-oriented adaptive per-flow path calculation mechanism in which source nodes assemble end-to-end routes with the assistance of group leaders at different hierarchy levels. Each leader is a DDQN agent operating on group-vector topology context and local resource metrics, and the reward is a deferred composite reward combining local resource assessments with delayed source-level end-to-end feedback. The resulting paths are described as directly applicable to segment routing, which places the method closer to hierarchical source-routed path computation than to classical hop-by-hop Q-routing (Ali et al., 2019).

In heterogeneous wireless networks, the same design pattern appears in a more explicitly local and distributed form. “Deep Reinforcement Learning for Multi-flow Routing in Heterogeneous Wireless Networks” formulates routing as a POMDP in which each node, using one shared pretrained model, jointly selects a next-hop relay, communication technology, and subband. The actual network is a dueling DQN with experience replay and $D^{(t)}$2-greedy exploration; the local state has dimension $D^{(t)}$3, and rate-based neighbor selection performs best in experiments. The method is evaluated under node mobility, changing density, and varying numbers of flows and resources, and the paper reports about 3% sum-rate reduction when testing in a 90-relay dense network after training with 45 or 63 relays, versus about 17% reduction when training with only 27 relays (Kim et al., 3 Nov 2025).

A closely related paper, “Deep Reinforcement Learning Based Routing for Heterogeneous Multi-Hop Wireless Networks,” also uses a dueling-DQN architecture with a shared policy across nodes, but optimizes the end-to-end bottleneck rate rather than sum-rate. The action is effectively a joint choice of next-hop relay and communication technology, realized by running the same DQN one resource at a time. The paper emphasizes candidate-neighbor design as a first-order issue: with $D^{(t)}$4 and $D^{(t)}$5, rate-based neighbor selection is best; with $D^{(t)}$6 and $D^{(t)}$7, the widest path-based algorithm achieves about $D^{(t)}$8 Mbps average while the DRL method with rate-based neighbor selection achieves $D^{(t)}$9 Mbps average, which the paper interprets as performance comparable to the optimal approach under favorable interference conditions (Kim et al., 20 Aug 2025).

Taken together, these systems show that DQRoute-style routing is not tied to a single action granularity. Depending on domain, the learned action can mean a request–path pair, a next hop, a next-hop–technology pair, or a hierarchy-level inter-group link. The common structure is a learned Q-function over feasible routing actions under dynamic resource constraints.

5. Adjacent quantum-routing baselines and the limits of shortest-path intuition

The non-learning quantum-routing literature provides an important baseline for interpreting DQRoute-style claims. In QKD networks, “Route Planning and Online Routing for Quantum Key Distribution Networks” argues that shortest-path routing performs poorly because the scarce resource is per-link key material rather than classical bandwidth. The paper formulates offline route planning as a path-based quadratic program with fairness objectives such as Prioritized Evenly Spread Contract Fulfillment (PESCF) and Evenly Distributed Granted Ratios (EDGR), and shows that in the online setting the widest shortest path strategy has a competitive ratio interpreted as at least $a \in \mathcal{A}$0, whereas shortest available path has poor worst-case behavior (López et al., 13 Aug 2025).

A newer decentralized entanglement-routing baseline, RADAR-Q, moves in a different direction from RL. It is a resource-aware distributed asynchronous routing protocol for multi-tenant quantum networks that embeds memory occupancy, link availability, and locality around a Nearest Common Ancestor (NCA) inside a DODAG hierarchy into path selection. In $a \in \mathcal{A}$1 grid and 100-node random-topology evaluations, RADAR-Q achieves aggregate throughputs $a \in \mathcal{A}$2 and $a \in \mathcal{A}$3 higher than synchronized and root-centric asynchronous baselines, respectively; under high load it maintains end-to-end fidelity above $a \in \mathcal{A}$4, reports Jain’s Fairness Index around $a \in \mathcal{A}$5–$a \in \mathcal{A}$6, and retains over $a \in \mathcal{A}$7 of ideal throughput at $a \in \mathcal{A}$8 ms coherence time (Tian et al., 29 Mar 2026).

These baselines matter because they discipline a frequent overstatement in DQRoute-style discussions: outperforming naive shortest path does not by itself establish superiority over the broader design space. In quantum networking, fidelity thresholds, buffer/key scarcity, root bottlenecks, and contention-aware locality already produce strong non-RL baselines. A plausible implication is that DQRoute-style systems are most informative when compared not only with shortest-path policies but also with contention-aware and theoretically analyzed alternatives.

6. DQRoute as “Divide, Weight, and Route” in long-tailed recognition

Outside routing, DQRoute is the exact acronym for a long-tailed recognition framework that combines difficulty-aware optimization with dynamic expert fusion. The method uses a shared backbone $a \in \mathcal{A}$9 and three parallel experts: a general expert $t$00, a medium-shot expert $t$01, and a tail expert $t$02. Each expert outputs both a class prediction $t$03 and an OOD/confidence score $t$04, and the inference-time routing weights are

$t$05

The final prediction is

$t$06

This is a decentralized expert-fusion mechanism: routing is performed by expert-specific confidence heads rather than by a separate centralized gate (Wei et al., 27 Aug 2025).

The “Weight” component is a class-reweighting scheme driven by class-wise difficulty. Difficulty is defined from average entropy $t$07 and exponentially smoothed class accuracy $t$08, and class weights are updated by

$t$09

These adaptive weights are then mixed with quantity-based class weights $t$10:

$t$11

Training is joint and end-to-end with a total objective

$t$12

The framework is evaluated on CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and Places-LT using ResNet-32, ResNeXt-50, and ResNet-152 backbones, SGD with momentum $t$13, weight decay $t$14, $t$15 epochs, and a learning rate linearly decayed from $t$16 (Wei et al., 27 Aug 2025).

The reported gains are concentrated on rare and difficult classes. On CIFAR-100-LT with imbalance ratio $t$17, DQRoute reports 51.7 overall accuracy with 61.9 many-shot, 52.7 medium-shot, and 38.6 few-shot, compared with 41.4 / 66.1 / 37.3 / 10.6 for CE and 49.4 / 61.6 / 50.5 / 33.9 for SADE. On Places-LT, DQRoute reports 41.3, ahead of CE at 31.4, RIDE at 40.3, and SADE at 40.9. This exact use of the name DQRoute is therefore unrelated to communication routing despite the shared vocabulary of “route” and “routing” (Wei et al., 27 Aug 2025).

Across these usages, DQRoute is best treated as an overloaded term rather than a single canonical method. In network routing, it denotes a family of deep Q-learning approaches whose clearest quantum-network instantiation in the provided literature is QuDQN. In visual recognition, it is the proper name of a mixture-of-experts framework. The two lines share a preference for adaptive decision rules over static heuristics, but they solve different optimization problems and should not be conflated.