Four papers. Four questions about making better use of wireless resources.
Choose a paper, follow its method stages, or select a concept to see how it connects to the rest of the study.
Concept maps explain the papers. Node positions, sizes and connections are schematic; they are not experimental data or a running research algorithm.
User-centric clustering
How can a learned model approximate joint user grouping and network-resource assignments?
Follow the path from a modeled wireless network to offline optimization labels and ensemble predictions. Keep the source study's evaluation separate from an operational feasibility guarantee.
Read the paper, methods & results ↗Follow the method
Identify channel information and the two coupled allocation decisions.
Source: Paper 1, Section II.A, PDF pp. 3-5.
The interactive map needs JavaScript and WebGL. Every concept, relationship and source is also available in the text view below.
Select a node or use the concept list. Drag to pan; use the zoom controls to inspect connections.
The same selection is available by keyboard.
Inputs & observations
A quasi-static user-centric network describes distributed base stations and user locations. The paper assumes TDD reciprocity and a modeled NOMA/SIC setting.
Source: Paper 1, Section II, PDF pp. 3-4.
The modeled geometry provides the context for channel information.
Inputs & observations
Estimated channel coefficients describe user-to-base-station links and inform the allocation model. This map does not compute or simulate those channels.
Source: Paper 1, Section II.A, PDF pp. 3-4.
The modeled geometry provides the context for channel information.
Channel conditions inform which users are grouped.
Channel conditions inform base-station and subcarrier assignments.
Channel coefficients join position information in the input vector.
Context & limits
The formulation includes membership, subcarrier, pilot and connectivity requirements. The source's receiver-sensitivity and assignment expressions require clarification before implementation.
Source: Paper 1, Eqs. (6)-(11), PDF pp. 4-5; editorial static review.
Allocation requirements accompany the objective transformation.
Decision or output
The x variables describe which cluster each user belongs to. This is distinct from assigning a base station and subcarrier to a cluster.
Source: Paper 1, Section II.A, PDF p. 4.
Channel conditions inform which users are grouped.
User membership is one part of the joint allocation representation.
Decision or output
The y variables describe base-station and subcarrier assignments for each cluster. The two output groups have different indexing structures.
Source: Paper 1, Section II.A, PDF p. 4; Section IV.B, PDF p. 8.
Channel conditions inform base-station and subcarrier assignments.
BS/subcarrier decisions form the other part of the joint representation.
Method
Auxiliary binary variables encode products of decisions; a first-order approximation replaces the rate term. An exact product transformation does not make the rate approximation exact or the integer problem polynomial-time.
Source: Paper 1, Section II.B, PDF pp. 5-6.
Allocation requirements accompany the objective transformation.
User membership is one part of the joint allocation representation.
BS/subcarrier decisions form the other part of the joint representation.
The transformed model is the input to the reported offline solver.
Method
CBC solves the study's linearized allocation model to generate supervision. The graph represents this reported role, not a validated implementation or a certificate for the original nonlinear problem.
Source: Paper 1, Section III, PDF p. 7.
The transformed model is the input to the reported offline solver.
The reported allocations become supervised targets.
Inputs & observations
Optimization-generated x and y decisions become targets for supervised learning. They are a model-based reference rather than measured network behavior.
Source: Paper 1, Sections III-IV, PDF pp. 7-8.
The reported allocations become supervised targets.
Offline labels guide model training; they are not new observations during inference.
Method
User positions, base-station positions and flattened channel coefficients form the feature vector. No fixed feature width or equal x/y split is inferred here.
Source: Paper 1, Section IV, PDF p. 8; presentation PDF p. 28.
Channel coefficients join position information in the input vector.
Normalized features are paired with the corresponding reference assignments.
Method
Multiple perceptrons learn the mapping using classification terms and constraint-inspired penalties. Predictions are averaged. Conflicting ensemble counts in the sources are not resolved by this visualization.
Source: Paper 1, Sections IV.A-IV.C, PDF p. 8; presentation PDF pp. 28-30.
Normalized features are paired with the corresponding reference assignments.
Offline labels guide model training; they are not new observations during inference.
Ensemble probabilities are averaged and rounded to candidate assignments.
Decision or output
The reported prediction stage averages probabilities and rounds them to binary assignments. Rounding alone does not certify membership, connectivity or resource feasibility.
Source: Paper 1, Sections IV.C-IV.E, PDF pp. 8-9.
Ensemble probabilities are averaged and rounded to candidate assignments.
The source evaluates prediction errors against optimization-generated references.
A separate feasibility audit is recommended; the paper does not establish it as an executed inference step.
Evaluation
The study compares predictions with optimization-generated targets using assignment accuracy and spectral-efficiency errors. The article page preserves the reported metrics and their units.
Source: Paper 1, Table II and Section IV.E, PDF p. 9.
The source evaluates prediction errors against optimization-generated references.
Context & limits
Before operational use, separately check predictions against an author-reviewed constraint model and define how failed checks are handled. This is an editorial recommendation, not a reported projection stage.
Source: Editorial recommendation; Paper 1, Section IV.E, PDF p. 9, states results without constraint projection.
A separate feasibility audit is recommended; the paper does not establish it as an executed inference step.
Identify channel information and the two coupled allocation decisions.
Source: Paper 1, Section II.A, PDF pp. 3-5.
Follow the reported offline transformation, solver and assignment labels.
Source: Paper 1, Sections II.B-III, PDF pp. 5-7.
Pair feature inputs with labels; penalties guide learning without guaranteeing feasibility.
Source: Paper 1, Sections IV.A-IV.C, PDF p. 8.
Read candidate decisions alongside the study evaluation and the separately recommended checks.
Source: Paper 1, Sections IV.C-IV.E, PDF pp. 8-9; editorial validation recommendation.
Pilot-efficient estimation
Can a learned refinement improve a classical channel estimate under a limited pilot budget?
Follow the LS-to-CNN inference path. A separate offline branch shows where ray-tracing data and GAN augmentation support model preparation.
Read the paper, methods & results ↗Follow the method
Known pilots and received observations define the estimation problem.
Source: Paper 2, Section II.A, PDF p. 2.
The interactive map needs JavaScript and WebGL. Every concept, relationship and source is also available in the text view below.
Select a node or use the concept list. Drag to pan; use the zoom controls to inspect connections.
The same selection is available by keyboard.
Inputs & observations
The receiver knows the pilot symbols. Pilot observations provide information about the channel but occupy time-frequency resources.
Source: Paper 2, Section II.A, PDF p. 2.
Pilot observations reflect the channel's transformation of a known signal.
Inputs & observations
The modeled received signal combines the known pilot matrix, channel and additive noise. These conceptual links are not a simulated waveform.
Source: Paper 2, Eq. (1), PDF p. 2.
Pilot observations reflect the channel's transformation of a known signal.
Received observations and known pilots support the coarse LS calculation.
Method
A regularized least-squares initialization supplies a coarse estimate for refinement. It avoids making sparsity a prerequisite of the proposed inference path.
Source: Paper 2, Section III.A, PDF p. 3, and Algorithm 1, PDF p. 4.
Received observations and known pilots support the coarse LS calculation.
The coarse estimate is transformed into the CNN input representation.
Method
The real and imaginary components are transformed using the specified preprocessing. Its associated inverse transformation is needed after the CNN.
Source: Paper 2, Algorithm 1, PDF p. 4; Section III.C, PDF p. 5.
The coarse estimate is transformed into the CNN input representation.
The normalized estimate is passed to the already trained CNN.
Method
An encoder-decoder CNN with skip connections refines the coarse representation. Its learned parameters are prepared offline.
Source: Paper 2, Fig. 1 and Section III.B, PDF pp. 3-4; presentation PDF p. 13.
The normalized estimate is passed to the already trained CNN.
CNN output is followed by the associated inverse preprocessing.
Offline training prepares the CNN. Neither training nor the GAN is an inference step.
Method
The associated inverse transformation returns the network output to a channel-estimate representation.
Source: Paper 2, Algorithm 1, PDF p. 4.
CNN output is followed by the associated inverse preprocessing.
Reverse normalization produces the final channel estimate.
Decision or output
The LS-to-CNN path returns the channel estimate. A smaller estimation error is not itself a measured end-to-end decoding reliability result.
Source: Paper 2, Algorithm 1 and Eq. (10), PDF p. 4.
Reverse normalization produces the final channel estimate.
The article examines error and resource use under its stated conditions.
Evaluation
The source compares NMSE, pilot budget, model cost and modeled spectral efficiency. Each addresses a different aspect of performance under the reported evaluation conditions.
Source: Paper 2, Section IV and Figs. 3-6, PDF pp. 5-6.
The article examines error and resource use under its stated conditions.
Inputs & observations
The offline data originates from the DeepMIMO outdoor scenario. The study describes sample selection and reconstruction-error filtering.
Source: Paper 2, Section III.C, PDF pp. 4-5.
Ray-tracing supplies the source channel realizations for offline training.
The GAN is described as learning from coarse estimates derived from source data.
Inputs & observations
Ray-traced reference channels supply supervised reconstruction targets. The study's sample-selection rules condition how generalization should be interpreted.
Source: Paper 2, Eqs. (8)-(9), PDF p. 4, and Section III.C, PDF p. 5.
Ray-tracing supplies the source channel realizations for offline training.
Known reference channels support the reconstruction loss for paired source examples.
Method
GAN augmentation is described as learning the distribution of coarse training estimates. It belongs to offline preparation. The paper does not fully specify the paired targets for every synthetic input.
Source: Paper 2, Section III.C and Eq. (11), PDF p. 5.
The GAN is described as learning from coarse estimates derived from source data.
Synthetic coarse examples support the described training strategy; their complete pairing protocol is not specified here.
Method
The CNN is trained with reconstruction loss and the described augmented dataset. This concept map does not invent a missing synthetic-input target-pairing mechanism.
Source: Paper 2, Sections III.B-III.C, PDF pp. 4-5.
Known reference channels support the reconstruction loss for paired source examples.
Synthetic coarse examples support the described training strategy; their complete pairing protocol is not specified here.
Offline training prepares the CNN. Neither training nor the GAN is an inference step.
Known pilots and received observations define the estimation problem.
Source: Paper 2, Section II.A, PDF p. 2.
Trace the inference path through LS, normalization, CNN and its inverse transformation.
Source: Paper 2, Algorithm 1, PDF p. 4.
See the training branch separately: ray-tracing, reference channels and GAN augmentation.
Source: Paper 2, Sections III.B-III.C, PDF pp. 4-5.
Keep error, pilot overhead and computation distinct, and retain the selected-data scope.
Source: Paper 2, Section IV, PDF pp. 5-6.
Energy & memory in IoT
How can storage roles adapt when both battery energy and available memory are limited?
CoMEM coordinates monitors and storage nodes within zones. Explore how data placement and role election balance resources, then examine the delay and processing trade-offs.
Read the paper, methods & results ↗Follow the method
Follow an event to its monitor and storage node, then inspect the query path.
Source: 2020 CoMEM article, storage model, PDF pp. 7-8.
The interactive map needs JavaScript and WebGL. Every concept, relationship and source is also available in the text view below.
Select a node or use the concept list. Drag to pan; use the zoom controls to inspect connections.
The same selection is available by keyboard.
Inputs & observations
Sensors generate event records in the WSN portion of an IoT system. Storage and forwarding consume memory and energy.
Source: 2020 CoMEM article, network model, PDF pp. 4-7.
The monitor coordinates the incoming event record.
Context & limits
Geographic hashing establishes initial monitor and storage destinations in the zoned data-centric organization.
Source: 2020 CoMEM article, storage model, PDF pp. 7-8.
The initial geographic mapping gives the zone a known monitor destination.
Method
The monitor coordinates storage and lookup, retains event summaries and maintains resource reports for its zone.
Source: 2020 CoMEM article, CoMEM description, PDF pp. 7-8.
The monitor coordinates the incoming event record.
The initial geographic mapping gives the zone a known monitor destination.
The monitor keeps summaries while details are sent to a storage node.
The monitor is the coordination point for an information request.
The replacement receives the outgoing monitor's records.
Extra coordination should be assessed alongside delay and processing work.
Decision or output
Storage nodes retain detailed records. Their selection can consider candidates throughout the zone rather than only a local neighbor set.
Source: 2020 CoMEM article, replica election, PDF pp. 8-9.
The monitor keeps summaries while details are sent to a storage node.
A query requiring details can be served using the stored records.
A selected candidate takes the storage/replica role.
Storage nodes apply the separate monitor-election energy test; this differs from joint-resource replica election.
Storage-role adaptation is evaluated for its data-loss behavior.
Inputs & observations
A query uses the monitor's summary or reaches the relevant storage node for detailed information.
Source: 2020 CoMEM article, storage and retrieval, PDF pp. 7-8.
The monitor is the coordination point for an information request.
A query requiring details can be served using the stored records.
Inputs & observations
Memory-related reports describe a resource that can limit data retention even when battery energy remains available.
Source: 2020 CoMEM article, resource definitions and performance table, PDF pp. 4-6 and 8.
Memory-related information is one dimension of replica election.
Inputs & observations
Energy information describes the ability to continue sensing, forwarding and coordinating. A resource-rich storage location can become unreachable if its relays are depleted.
Source: 2020 CoMEM article, motivation and resource model, PDF pp. 4-7.
Energy-related information is considered together with memory.
Residual-energy information supports a separate monitor-change decision.
Method
The replica-election heuristic compares energy and memory coefficients and coverage counts when selecting storage nodes. It is a practical selection rule, not a proof of a global optimum.
Source: 2020 CoMEM article, Eqs. (23)-(25) and Fig. 4, PDF pp. 8-9.
Memory-related information is one dimension of replica election.
Energy-related information is considered together with memory.
The joint resource goal motivates the heuristic election rule.
A selected candidate takes the storage/replica role.
Decision or output
Storage nodes apply a separate energy-gap test that can trigger a new monitor. Stored monitor records transfer to the replacement so that coordination need not remain on one depleted node.
Source: 2020 CoMEM article, Eq. (26) and monitor election, PDF p. 9.
Residual-energy information supports a separate monitor-change decision.
Storage nodes apply the separate monitor-election energy test; this differs from joint-resource replica election.
The replacement receives the outgoing monitor's records.
Evaluation
The study evaluates data loss alongside resource use. The graph describes the objective of keeping records available, not a guarantee of zero loss.
Source: 2020 CoMEM article, performance evaluation and Table 3, PDF pp. 11-15.
Storage-role adaptation is evaluated for its data-loss behavior.
Evaluation
Resource coordination has a cost. The published ranking places basic DCS ahead of CoMEM on delay and computational work.
Source: 2020 CoMEM article, Table 3, PDF pp. 14-15.
Extra coordination should be assessed alongside delay and processing work.
Context & limits
The cross-layer formulation combines memory, energy and load balancing. CoMEM is a heuristic for this coupled resource problem.
Source: 2020 CoMEM article, Eqs. (1)-(22), PDF pp. 4-7.
The joint resource goal motivates the heuristic election rule.
Follow an event to its monitor and storage node, then inspect the query path.
Source: 2020 CoMEM article, storage model, PDF pp. 7-8.
A node may have energy but little memory, or memory but little energy.
Source: 2020 CoMEM article, resource definitions, PDF pp. 4-6.
Replica election combines resources; monitor rotation is a separate energy-based process.
Source: 2020 CoMEM article, Fig. 4 and Eqs. (23)-(26), PDF pp. 8-9.
Interpret data availability together with coordination delay and processing cost.
Source: 2020 CoMEM article, Table 3, PDF pp. 14-15.
Adaptive sensor storage
When should a sensor network split or merge its storage zones as event traffic changes?
AR-DCS adapts a zoned storage layout to incoming events. Smaller zones can shorten storage paths, while more replicas increase the work required for queries.
Read the paper, methods & results ↗Follow the method
Monitor event arrivals using a threshold grounded in the network model.
Source: 2016 AR-DCS article, Eqs. (1)-(4), PDF pp. 4-5.
The interactive map needs JavaScript and WebGL. Every concept, relationship and source is also available in the text view below.
Select a node or use the concept list. Drag to pan; use the zoom controls to inspect connections.
The same selection is available by keyboard.
Inputs & observations
Storage packets arrive at zone monitors. Their frequency is the observation driving the proposed layout adaptation.
Source: 2016 AR-DCS article, method description, PDF p. 4.
The monitor observes the arriving storage traffic.
Inputs & observations
Sensing range, node separation, target count and zone count inform the proposed threshold. These are model assumptions, not universal constants.
Source: 2016 AR-DCS article, Table 1 and Eqs. (1)-(4), PDF pp. 4-5.
Geometric and density assumptions determine the proposed threshold inputs.
Method
The threshold combines the modeled sensing count and density with targets and current zones. This graph does not calculate or select a deployment threshold.
Source: 2016 AR-DCS article, Eq. (4), PDF p. 5.
Geometric and density assumptions determine the proposed threshold inputs.
Observed demand is assessed against the proposed threshold.
Method
Monitors sample storage arrivals and evaluate adaptation at the cadence described in the paper, separating observation from repeated reconfiguration.
Source: 2016 AR-DCS article, implementation considerations, PDF p. 4.
The monitor observes the arriving storage traffic.
Observed demand is assessed against the proposed threshold.
Monitors use the monitoring map to coordinate layout changes.
Context & limits
The monitoring map supports coordination among monitors when the storage layout changes.
Source: 2016 AR-DCS article, zone coordination, PDF pp. 4 and 7.
Monitors use the monitoring map to coordinate layout changes.
The division decision responds to the relevant high-arrival condition.
Aggregation responds to a return to lower event demand.
Decision or output
A busy parent zone divides into four subzones with new monitors and replicas. The changed layout distributes storage traffic across shorter paths.
Source: 2016 AR-DCS article, division mechanism and Figs. 1-3, PDF pp. 4 and 7.
The division decision responds to the relevant high-arrival condition.
Subzones receive monitor and replica roles under the changed mapping.
The storage benefit is evaluated in the source's traffic comparison.
Decision or output
When event demand falls, subzones can merge into their parent. Fewer monitor and replica destinations reduce the query-side burden.
Source: 2016 AR-DCS article, aggregation discussion, PDF pp. 4 and 7.
Aggregation responds to a return to lower event demand.
Aggregation restores a coarser organization of monitor and replica destinations.
Old subzone data remains available during the described query transition.
Method
The current layout maps event data to monitor and replica locations. Replication remains part of the design when zone granularity changes.
Source: 2016 AR-DCS article, DCS foundations and AR-DCS, PDF pp. 3-5.
Subzones receive monitor and replica roles under the changed mapping.
Aggregation restores a coarser organization of monitor and replica destinations.
A larger replica/monitor set may increase query and reply traffic.
Context & limits
Old subzone replicas retain data for a transition period during aggregation so in-flight queries can still be answered. A universal retention duration is not specified.
Source: 2016 AR-DCS article, query handling during aggregation, PDF p. 4.
Old subzone data remains available during the described query transition.
Still-valid transition replicas preserve a path to previously stored data.
Evaluation
Shorter paths can reduce storage forwarding as demand increases. The source reports the trend in its scenario comparison.
Source: 2016 AR-DCS article, results discussion and Figs. 4-7, PDF pp. 7-9.
The storage benefit is evaluated in the source's traffic comparison.
Storage traffic is only one part of the overall traffic comparison.
Evaluation
A divided layout can require queries and responses from more monitors and replicas. This cost must be read with the storage benefit.
Source: 2016 AR-DCS article, results discussion and Figs. 4-7, PDF pp. 8-9.
A larger replica/monitor set may increase query and reply traffic.
Still-valid transition replicas preserve a path to previously stored data.
Query and response overhead must be included in the same interpretation.
Evaluation
The paper compares storage, query, response and total traffic under matched scenario conditions. It reports overall traffic similar to basic DCS while retaining replication.
Source: 2016 AR-DCS article, evaluation protocol and results, PDF pp. 6-9.
Storage traffic is only one part of the overall traffic comparison.
Query and response overhead must be included in the same interpretation.
Monitor event arrivals using a threshold grounded in the network model.
Source: 2016 AR-DCS article, Eqs. (1)-(4), PDF pp. 4-5.
Coordinate splitting at higher demand and merging as demand falls.
Source: 2016 AR-DCS article, method and zone layouts, PDF pp. 4 and 7.
The current mapping and transition replicas support queries during change.
Source: 2016 AR-DCS article, query-transition discussion, PDF p. 4.
Read storage-path savings with the added query and response traffic.
Source: 2016 AR-DCS article, Figs. 4-7 and discussion, PDF pp. 7-9.
Concept maps use Sigma.js and Graphology, served locally. Open-source notices.
