CRouting: Adaptive Routing Under Uncertainty

• CRouting is a family of routing formalisms addressing transmission path selection under uncertainty and constrained resources across diverse domains.
• Key studies leverage probabilistic modeling, hierarchical clustering, and prediction-based topology control to improve reliability and efficiency.
• Empirical evaluations show CRouting methods reduce overhead and enhance route stability compared to traditional routing under dynamic conditions.

CRouting denotes several routing problem classes in the surveyed literature, all centered on selecting an effective transmission, forwarding, coordination, or execution path under uncertainty and constrained resources. In the largest cluster of work, it refers to routing in cognitive radio networks, where secondary users must exploit spectrum holes while respecting primary-user priority, mobility, and heterogeneous channel availability (Benidris et al., 2016, Guan et al., 2011, Zhong et al., 2015, Dall'Anese et al., 2012). Other papers use the term for carrier selection in VoIP termination, coordination-group formation for en-route travelers, backhaul-aware message placement in cloud radio access networks, and local-versus-edge model selection for wireless LLM inference (0912.4241, Peng et al., 2019, Ahmad et al., 2019, Xue et al., 12 May 2026). This suggests that CRouting is best viewed as a family of routing formalisms rather than a single protocol.

1. Scope and terminology

Across the surveyed set, the term spans both classical network-layer routing and broader resource-allocation problems that are treated as routing because they decide where traffic, messages, or queries should go. The cognitive-radio literature is closest to the conventional networking meaning: route discovery, relay selection, topology control, opportunistic forwarding, and route maintenance over multi-hop wireless links whose usability depends on primary-user activity (Benidris et al., 2016, Guan et al., 2011, Zhong et al., 2015, Dall'Anese et al., 2012). By contrast, the C-RAN and device-edge literatures interpret routing more abstractly as message placement or query deferral across cooperating service tiers (Ahmad et al., 2019, Xue et al., 12 May 2026).

Domain	Routing object	Representative work
Cognitive radio networks	SU/CU forwarding under PU constraints	(Benidris et al., 2016, Guan et al., 2011, Zhong et al., 2015, Dall'Anese et al., 2012)
VoIP carrier routing	Vendor selection for call termination	(0912.4241)
Coordinated road traffic routing	Traveler grouping and coordinated route choice	(Peng et al., 2019)
C-RAN	Private/common stream placement across BS clusters	(Ahmad et al., 2019)
Device-edge LLM serving	Local accept/defer and edge-model selection	(Xue et al., 12 May 2026)

A recurring distinction is whether routing acts on packets over a graph, on users over a road network, or on computational requests over a service hierarchy. The surveyed literature nevertheless shares a common structure: routing quality depends on latent constraints that are not captured by shortest-path reasoning alone.

2. Cognitive-radio routing as the canonical meaning

In cognitive radio ad hoc networks and CR-MANETs, routing is difficult because forwarding decisions depend simultaneously on topology dynamics and spectrum dynamics. Secondary users can access licensed spectrum only when primary users are inactive, so route viability changes with white-space availability, PU activity, mobility, and the fact that neighboring nodes may not share the same available channels (Benidris et al., 2016). One paper states the problem directly: route discovery in CRAHNs must reduce control-message flooding, maintain routes under mobility, and select forwarding nodes compatible with local channel availability (Benidris et al., 2016).

The CR-MANET formulation sharpens this by redefining link existence itself. A CU-CU link is usable only when the two cognitive users remain within transmission range and neither endpoint moves into the interference region of any active PU (Guan et al., 2011). In that setting, shortest-path routing based only on hop count can be unstable, because a geometrically short path may traverse a node moving toward a PU exclusion region, whereas a slightly longer path may be substantially more durable (Guan et al., 2011).

A further generalization appears in statistical routing, where the route is not a fixed deterministic path at all. Instead, node $U_n$ forwards to neighbor $U_i$ with probability $t_{n\rightarrow i}$, transmission succeeds with probability $r_{n\rightarrow i}$, and the effective per-link forwarding mechanism is the product $t_{n\rightarrow i} r_{n\rightarrow i}$ (Dall'Anese et al., 2012). This replaces path selection over a static graph by probabilistic flow allocation over a stochastic graph whose edge quality depends on fading, interference, random access collisions, and PU protection constraints (Dall'Anese et al., 2012).

The surveyed set therefore treats cognitive-radio CRouting as a cross-layer problem. Route quality depends on whether a link remains usable long enough, whether adjacent nodes share a common channel, whether PU interference limits future transmissions, and whether MAC- and PHY-layer decisions make the route statistically sustainable.

3. Protocol families and routing metrics in cognitive-radio CRouting

One major family is hierarchical clustering. The cluster-based CRAHN protocol organizes nodes into clusters, elects a cluster head based on being a static node, having superior energy resources, and having long lifetime, and then uses the cluster head for neighbor-table maintenance, routing-table maintenance, route discovery, and route repair (Benidris et al., 2016). Route discovery begins with $S \to C_s$ using RREQ messages identified by request ID, source IP address, and destination IP address; duplicate RREQs are discarded; discovery is bounded by a maximum hop threshold $H_{max}$; and the destination cluster head starts a timer $T_R$ upon receipt of the first RREQ (Benidris et al., 2016). The implemented route-selection rule is two-level: select the inter-cluster path with minimal hop-count, then choose within each intermediate cluster a CR node that shares a common channel with the previous hop and has maximum throughput (Benidris et al., 2016).

A second family inserts prediction before routing. Prediction-based Cognitive Topology Control (PCTC) is a middleware-like cross-layer module between the CR module and routing, intended to let protocols such as AODV or DSR inherit cognitive behavior without full redesign (Guan et al., 2011). Its central object is the cognitive link availability

$$T_a = \min \Big\{ T_p L(T_p),\ \tilde{T}_{p_i}^{\,j} L(\tilde{T}_{p_i}^{\,j}) \Big\},$$

which combines mobility-induced link lifetime, PU-interference-limited lifetime, and prediction confidence (Guan et al., 2011). PCTC then assigns each link the weight

$w = r(T_a-\delta),$

and each path the bottleneck weight

$U_i$0

so that topology control preserves the reliable path with maximum path weight while reducing node degree and contention (Guan et al., 2011).

A third family is opportunistic and coding-aware. CROR combines opportunistic routing, inter-session network coding, and the Successful Delivery Ratio metric SuDR (Zhong et al., 2015). For packet size $U_i$1 and channel rate $U_i$2, the transmission time is

$U_i$3

and the probability that channel $U_i$4 remains available long enough is

$U_i$5

SuDR then combines successful spectrum utilization probability and packet loss rate, and the multi-hop form is recursively defined so that forwarding success is the complement of the event that all candidate relays fail (Zhong et al., 2015). CROR further biases forwarding timers toward relays with both high SuDR and many coding opportunities: $U_i$6 This places spectrum availability, loss, and coding gain inside the same forwarding rule (Zhong et al., 2015).

A fourth family formalizes routing as cross-layer optimization. In statistical routing for cognitive random-access networks, link success is approximated as

$U_i$7

and routing, transmission probabilities, and transmit powers are jointly optimized under queue-stability and PU interference constraints (Dall'Anese et al., 2012). This is not route computation over fixed edges; it is probabilistic flow design driven by outage statistics.

4. Evaluation trends, critiques, and recurring limitations

The empirical literature is consistent in reporting that cognitive-radio-aware routing outperforms flat or non-cognitive baselines when the source of uncertainty is modeled explicitly. The cluster-based CRAHN protocol is reported to achieve higher route discovery success, lower RREQ overhead, lower RREP overhead, lower routing delay, and better scalability with increasing CR node count than AODV, although the paper gives trends rather than exact numeric tables (Benidris et al., 2016). PCTC reduces average and maximum node degree, retains longer-lived links, and improves throughput and delay relative to routing on the unpruned predicted topology, especially because lower contention remains beneficial even when the same reliable path is preserved by the 1-spanner property (Guan et al., 2011). CROR is reported to obtain better results than ExOR, MORE, and MaxPoS with respect to throughput, the probability of PU-SU packet collision, and spectrum utilization efficiency (Zhong et al., 2015). Statistical routing shows route shifts away from PU-affected regions and a preference for non-shortest paths when those are statistically more reliable under power and interference constraints (Dall'Anese et al., 2012).

At the same time, the literature contains direct critiques of what counts as routing. The comment on CRB-RPL argues that the protocol is largely a repackaging of CRB-MAC receiver-based forwarding, that the DAG structure of RPL is not retained, and that rank or CTQ alone does not guarantee primary receiver protection because the sender does not explicitly enforce forwarder priority (Aijaz, 2017). The same critique states that rank difference is not hop count, that some analytical parameters are undefined or dimensionally incorrect, and that the coordination-overhead model is wrong because it neglects cases in which a node misses either the preamble or the data frame and therefore forwards redundantly (Aijaz, 2017).

Several limitations recur across the cognitive-radio papers. The cluster-based CRAHN protocol explicitly leaves the value of $U_i$8, the distance formula, and the cluster assignment objective unspecified, making the clustering heuristic rather than fully specified (Benidris et al., 2016). PCTC treats PU interference boundaries geometrically and gives limited treatment to PU activity dynamics, multi-channel selection, and channel heterogeneity (Guan et al., 2011). CROR assumes exponential ON/OFF PU behavior and a two-radio architecture, and it does not deeply quantify signaling overhead or coding-table maintenance cost (Zhong et al., 2015). Statistical routing depends on channel and interference statistics and reaches a KKT point of a non-convex problem rather than a guaranteed global optimum (Dall'Anese et al., 2012). A plausible implication is that CRouting research often achieves practical adaptability before complete modeling closure.

5. CRouting beyond cognitive radio

In telephony, CRouting denotes carrier routing under untrusted signaling. The dynamic-routing system for ToIP termination addresses false answer supervision, where an upstream vendor returns SIP 2xx even though the call is not actually connected, and therefore backup routing is never triggered by the softswitch (0912.4241). The implemented solution inserts a Kamailio layer between billing and the real vendors, computes per-vendor ACD over dynamic intervals of at least 20 minutes and 20 calls, preserves a minimum exploration load $U_i$9, and probabilistically returns 480 Temporarily Unavailable to force retry through the next vendor (0912.4241). In a worked example with ACD values $t_{n\rightarrow i}$0 min and $t_{n\rightarrow i}$1 min, the resulting target balance is approximately $t_{n\rightarrow i}$2 versus $t_{n\rightarrow i}$3, showing how route quality feedback can be translated into runtime route suppression (0912.4241).

In online coordinated road routing, the routing object is the set of travelers rather than packets. The key construct is competition potential, interpreted by the paper as the potential that travelers will use the same traffic roads in a future time interval (Peng et al., 2019). Direct and indirect coordination potential are embedded in a competition network, then the Adaptive Centroid-based Clustering Algorithm (ACCA) partitions travelers into coordination groups, and the clustering-based coordinated routing mechanism (CB-CRM) applies the earlier CRM inside each group (Peng et al., 2019). The main reported result is computational: CB-CRM significantly improves computation efficiency with minor system performance loss in large networks, and the advantage becomes more apparent under high penetration and congested traffic condition (Peng et al., 2019).

In C-RAN, routing is reframed as cloud-to-edge message placement. The RS-CMD formulation splits each user message into private and common parts, decides which BSs serve each part, and jointly optimizes clustering and beamforming under per-BS backhaul capacity and transmit power constraints (Ahmad et al., 2019). The BS-side sets $t_{n\rightarrow i}$4 and $t_{n\rightarrow i}$5 function as stream-placement tables, and the backhaul constraint makes wider dissemination of a stream expensive even when it improves coordinated interference mitigation (Ahmad et al., 2019). The paper explicitly notes that this is not packet routing in the network-layer sense, but a message-routing or placement problem inside the cloud radio infrastructure (Ahmad et al., 2019).

6. Device-edge LLM routing and the contemporary generalization of CRouting

The most recent surveyed formulation treats CRouting as query-level routing for wireless device-edge LLM inference. CR$t_{n\rightarrow i}$6 assumes a two-tier architecture with a lightweight UE model $t_{n\rightarrow i}$7 and a pool of stronger edge models $t_{n\rightarrow i}$8, and asks whether a query should be answered locally or deferred to the edge under latency, energy, and accuracy tradeoffs (Xue et al., 12 May 2026). Because the UE cannot observe edge-model utilities before deferral, the deployable policy is factorized: $t_{n\rightarrow i}$9 The utility for model $r_{n\rightarrow i}$0 is

$r_{n\rightarrow i}$1

and the on-device gate learns the teacher-induced local-versus-edge margin while the edge-side selector chooses

$r_{n\rightarrow i}$2

for deferred queries (Xue et al., 12 May 2026).

The distinctive contribution is conformal risk control. For fixed operating point $r_{n\rightarrow i}$3 and risk level $r_{n\rightarrow i}$4, the calibrated threshold

$r_{n\rightarrow i}$5

controls marginal false-acceptance risk, where false acceptance means that the UE answers locally even though the full-information teacher utility prefers an edge model (Xue et al., 12 May 2026). Empirically, CR$r_{n\rightarrow i}$6 is reported to reduce normalized deployment cost by up to $r_{n\rightarrow i}$7 at matched accuracy relative to strong query-level baselines, while the device-edge router head requires $r_{n\rightarrow i}$8 parameters and $r_{n\rightarrow i}$9k FLOPs excluding the shared encoder (Xue et al., 12 May 2026).

This suggests a broad contemporary generalization of CRouting. In the cognitive-radio literature, the central question is whether a packet can traverse a link before topology or spectrum conditions invalidate it. In the device-edge LLM literature, the analogous question is whether a query should traverse the wireless edge boundary before cost and risk dominate the expected utility of local execution. In both cases, routing quality is determined less by nominal reachability than by constrained observability, dynamic resource costs, and the asymmetry of wrong decisions.