PaperlayerView exact revision on arXiv

Source-linked AI summary

Concurrent Task Orchestration: Hierarchical Trusted Collaborator Group Formation for Constrained Resource Allocation

Jiazhi Chen, Xianbin Wang, Xuemin "Sherman" Shen

arXiv:2610.33295v1cs.NI

TL;DR

Concurrent tasks, limited resources, and changing collaborator availability complicate meeting task-specific completion goals. The paper combines trusted group formation, hierarchical matching, task-specific trust prediction, and fairness-aware learning; simulations report a 22% improvement over representative baselines.

  • Problem

    Concurrent task demands, limited distributed resources, and dynamic collaborator availability make it difficult to satisfy task-specific completion goals in collaborative systems.

    Source: §I · p. 1, A. Motivation · p. 2

  • Method

    A unified value-risk utility model supports α-fair long-term orchestration using hierarchical task-collaborator matching, spatial-temporal GNN trust prediction, and fairness-aware DRL.

    Source: C. Contributions · p. 3, §V · p. 12

  • Results

    22% improvement in concurrent task completion performance over representative baselines under resource constraints.

    Source: §V · p. 12

  • Takeaways & Limitations

    The proposed strategy also maintains an average attack resistance rate of approximately 93% when 50% of UGVs launch on-off attacks.

    Source: §V · p. 12

Abstract

from arXiv · show

Effective collaboration among resource-constrained mobile intelligent machines, such as Unmanned Aerial Vehicles (UAVs), is essential for the completion of complex tasks. However, concurrent task demands, limited distributed resources, and dynamic collaborator availabilities make it challenging to satisfy task-specific completion goals in diverse collaborative systems. To address these challenges simultaneously, we propose a new framework of trusted collaborator group formation and the supporting resource allocation strategy for concurrent task orchestration. Specifically, a collaborative task is decomposed into subtasks subject to dynamic resource availability and inter-subtask dependencies. A task utility model is developed for collaborator group evaluation support subtask allocation according to task completion goals and unified value-risk analysis. Based on this model, concurrent task orchestration is formulated as a long-term utility maximization problem under $\alpha$-fairness. To solve the problem efficiently, we develop a hierarchical matching strategy that first clusters tasks and collaborators into coarse-grained sets and subsequently performs fine-grained collaborator group formation and task allocation within matched sets. Task-specific trust is predicted through spatial-temporal Graph Neural Networks (GNNs) and embedded into a fairness-aware Deep Reinforcement Learning (DRL) framework to continuously update the matching policies. Simulation results demonstrate that the proposed strategy improves concurrent task completion performance by 22% over representative baselines under resource constraints.

I. INTRODUCTION

Concurrent task orchestration must reconcile diverse task goals, changing collaborator conditions, limited resources, and long-term fairness. The paper proposes trusted group formation and task allocation using hierarchical matching, task-specific trust, and fairness-aware learning.

A. Motivation

  • Resource constraints and dynamic collaborators make it difficult to meet diverse completion goals while coordinating concurrent tasks and maintaining long-term fairness.The paper frames effective orchestration as jointly accounting for heterogeneous goals, changing collaborator conditions, scalable group formation, and fairness.Source: A. Motivation · pp. 1–2
  • Task decomposition and collaborator selection jointly shape task value and risk, while static-information heuristics are ill-suited to highly dynamic scenarios.The discussion highlights latency and energy goals, attacks, uncertainty in collaborator conditions, and the need to preserve subtask dependencies.Source: A. Motivation · p. 1
  • Maximizing cumulative utility alone can leave some concurrent tasks persistently underserved, motivating long-term fairness in resource allocation.Short-term fairness may be unattainable under resource constraints, but the paper identifies long-term fairness as a way to prevent persistent resource monopolization.Source: A. Motivation · p. 2

B. Related Work

3) Fairness-Aware Task Orchestration:

  • Prior fairness-aware orchestration often targets specific fairness criteria or relies on heuristic or Monte Carlo methods, rather than integrating tunable α-fairness with trust-aware hierarchical decisions.The cited approaches are described as decoupled from trust-aware and hierarchical decision-making.Source: 3) Fairness-Aware Task Orchestration · pp. 2–3

C. Contributions

  • The proposed strategy jointly forms trusted collaborator groups and allocates tasks under dynamic, resource-constrained conditions.Its utility model captures task goals and deviation through value-risk analysis, and it formulates orchestration as long-term average utility maximization under α-fairness.Source: C. Contributions · p. 3
  • Tasks and collaborators are clustered before fine-grained group formation and allocation, with spatial-temporal GNN trust representations guiding fairness-aware DRL policies.The matching process uses predicted trust as abstracted states and fairness-aware rewards adjusted using task utilities and historical reward statistics.Source: C. Contributions · p. 3

A. Modeling Tasks and Resources as Graphs

Tasks and UGV resources are represented as graphs to preserve dependencies and communication structure. Completion value accounts for task goals, including time, whose overall task measure is set by its slowest subtask.

A. Modeling Tasks and Resources as Graphs

  • Each task is an undirected graph of subtasks, dependencies, and required contact durations, while UGV resources form a graph of one-hop V2V links and VMs.The resource graph also records exponential contact-duration rate parameters, and UGV VMs have differing CPU capabilities.Source: A. Modeling Tasks and Resources as… · p. 3

B. Modeling Value of Task Completion

  • Value of Task Completion evaluates how effectively a collaborator group meets task goals using time, energy, and reliability values.The task-level completion time is determined by the slowest subtask; completion goals also include transmission, computation, and dependent-subtask V2V exchange delays.Source: B. Modeling Value of Task Completion · pp. 3–4, 1) Value of Time · p. 4

1) Value of Time:

  • Subtask completion time reflects data size, required CPU cycles, available VM frequency, UAV-to-UGV transmission rate, and dependent-subtask data-exchange delay.Overall task completion time is the maximum subtask completion time.Source: 1) Value of Time · p. 4

2) Value of Energy:

The task-completion value model evaluates energy and reliability alongside time. Energy accounts for offloading transmission and channel holding, while reliability combines authentication and transmission secrecy.

2) Value of Energy:

  • UAV offloading energy includes transmission and channel-holding energy, and subtask energy value compares expected energy consumption with actual consumption.The overall energy consumption for a task aggregates the energy consumed across its subtasks.Source: 2) Value of Energy · p. 4

3) Value of Reliability:

  • Collaborator reliability combines identity authentication and transmission secrecy, with reliability value adapted to each subtask’s security requirements.Authentication uses locally executed RSU checks; the paper also describes false-alarm and misdetection rates in determining authentication results.Source: 3) Value of Reliability · pp. 4–5

C. Modeling Risk of Task Completion

The formulation combines task-specific completion risk with task utility, then maximizes long-term utility under α-fairness and resource constraints. Its mixed-integer, uncertain, long-term objective motivates a hierarchical solution.

C. Modeling Risk of Task Completion

  • Task-specific trust thresholds define risk around dynamic completion goals, and overall task risk includes both collaborator assignment risk and compounded inter-subtask dependency risk.The risk model evaluates deviation from completion goals using normalized performance measures and weights reflecting subtask importance.Source: C. Modeling Risk of Task Completion · p. 5

D. Problem Formulation

  • Task utility combines obtained value and perceived risk, and the objective maximizes long-term concurrent-task utility under α-fairness and allocation constraints.The constraints limit each subtask to one UGV, respect available VMs, require sufficient V2V contact probability for dependent subtasks assigned to different UGVs, and restrict offloading to communication range.Source: D. Problem Formulation · pp. 5–6
  • The mixed-integer problem is difficult because uncertain dynamic conditions complicate long-term optimization while resource contention requires balancing total utility and long-term fairness.The paper notes that short-term fairness variations may be needed for urgency and limited-resource conditions.Source: D. Problem Formulation · p. 6

III. HIERARCHICAL TRUSTED COLLABORATOR GROUP FORMATION AND TASK ALLOCATION

The framework predicts task-specific trust and uses it to form collaborator groups and allocate tasks through coarse-to-fine matching. It first clusters tasks and resources by feature similarity, then predicts trust across and within matched sets.

III. HIERARCHICAL TRUSTED COLLABORATOR GROUP FORMATION AND TASK ALLOCATION

  • The proposed strategy combines hierarchical task-specific trust prediction via GNNs with trust-embedded fairness-aware DRL for task allocation under long-term fairness.Online inference is coordinated by nearby RSUs; training is performed offline and periodically.Source: §III · p. 6

A. Hierarchical Task-Specific Trust Prediction via GNNs

  • Task-specific trust abstracts the value and risk of task-resource interactions that arise from dynamic task requirements and collaborator conditions.HTST-GNN dynamically coarsens tasks and resources and predicts trust at both coarse and fine levels.Source: A. Hierarchical Task-Specific Trust… · p. 6
  • Trust is predicted for time t+1 from a sequence of trust-guided bipartite graphs, first between task and resource sets and then within matched sets.The graphs represent time-varying subtask, collaborator, and trust features.Source: A. Hierarchical Task-Specific Trust… · p. 6

1) Dynamic Coarsening Process:

  • Dynamic coarsening partitions feature-similar tasks to reduce task-collaborator mismatches and the subsequent matching space.K-means clustering produces the task and resource sets used to form a coarsened trust-guided bipartite graph.Source: 1) Dynamic Coarsening Process · p. 6
  • Concurrent-task coarsening clusters subtasks by feature similarity, with clusters potentially combining subtasks from different UAVs and retaining their task dependencies.Cluster centroids are formed from the features of their subtasks.Source: 1) Dynamic Coarsening Process · p. 6
  • Resource coarsening clusters collaborators by feature similarity, reflecting their heterogeneous identities, secrecy capabilities, and computing resources.Each resource cluster comprises collaborators and their available VMs.Source: 1) Dynamic Coarsening Process · p. 6

2) Spatial-Temporal Feature Aggregation:

Spatial-temporal aggregation builds common-space embeddings for coarsened task and resource graphs, updates them using neighboring information, and attends to historical embeddings. The procedure stores coarse- and fine-grained trust predictions.

2) Spatial-Temporal Feature Aggregation:

  • Raw cluster-centroid features for task sets, collaborator sets, and their edges are transformed into a common vector space for spatial-temporal aggregation.The framework overview identifies task-specific trust prediction via spatial-temporal GNNs as one of its two modules.Source: 2) Spatial-Temporal Feature Aggregation · p. 6, Figure 2 · p. 7
  • At each layer, task-set centroids aggregate information propagated from neighboring collaborators, using edge-aware weighted sums and GRUs for updates.Resource-set centroids are updated symmetrically.Source: 2) Spatial-Temporal Feature Aggregation · pp. 6–7
  • Historical embeddings are combined through attention, with normalized coefficients measuring each time slot’s relative importance to the target slot.Learnable matrices parameterize the attention mechanism.Source: 2) Spatial-Temporal Feature Aggregation · p. 7
  • The HTST-GNN procedure stores coarse- and fine-grained trust matrices after predicting trust between task and resource sets and within matched sets.Coarse trust is based on aggregated edge trust values, including an initial trust estimate that balances task attributes.Source: 2) Spatial-Temporal Feature Aggregation · p. 7

3) Prediction and Training:

The prediction stage maps learned embeddings to edge representations and trust scores, while restricting training to similar tasks. It trains the trust model with cross-entropy and reuses or fine-tunes parameters according to task similarity.

3) Prediction and Training:

  • An MLP maps final embeddings to edge embeddings, which are transformed into trust values in the range [0, 1] by a sigmoid function.The trust mapping uses learnable parameter matrices.Source: 3) Prediction and Training · p. 7
  • Task-specific training prunes the graph to task sets sufficiently similar to the new subtask, with similarity measured by cosine similarity against prior cluster centroids.The similarity threshold determines which task sets are retained.Source: 3) Prediction and Training · p. 8
  • The trust model is optimized with cross-entropy loss and Adam; parameters are reused for highly similar tasks and fine-tuned otherwise.The procedure applies spatial-temporal aggregation and task-specific training to candidate collaborators from matched resource sets.Source: 3) Prediction and Training · p. 8
  • After coarse-grained trust prediction selects candidate resource sets, fine-grained prediction estimates trust between their subtasks and collaborators.The resulting fine-grained trust predictions support collaborator-group formation within the candidate pools.Source: 3) Prediction and Training · p. 8

B. Trust-Embedded Fairness-Aware DRL

The fairness-aware DRL framework reduces matching complexity through coarse task-set/resource-set matching followed by fine subtask/collaborator matching. Its coarse policy uses predicted trust as state information and adjusts rewards to prioritize tasks with lower long-term utility.

B. Trust-Embedded Fairness-Aware DRL

  • Hierarchical DRL progressively reduces the matching space: a coarse policy matches task and resource sets, then a fine policy matches subtasks and collaborators within selected sets.The fine policy is conditioned on the coarse matching policy.Source: B. Trust-Embedded Fairness-Aware DRL · p. 8
  • The coarse-stage objective aggregates subtask value and risk into long-term task-set utility, while allowing many-to-many set matching subject to available-VM limits.The fine stage optimizes long-term utility at the subtask level within selected sets.Source: B. Trust-Embedded Fairness-Aware DRL · p. 8

1) Stage I–Design and Learning of Coarse-Grained Level Matching Policy:

  • Task sets act as learning agents in an infinite-horizon discounted MDP, using predicted task-specific trust as the coarse-grained state and set-matching decisions as actions.The policy is represented by an artificial neural network.Source: 1) Stage I–Design and Learning of… · p. 8
  • The coarse-stage instantaneous reward is task-set utility weighted by the derivative of the fairness utility, which assigns higher weights to tasks with lower long-term utility.The running average is used because long-term utility is unavailable online.Source: 1) Stage I–Design and Learning of… · p. 8
  • An actor-critic DRL algorithm optimizes the α-fair utility, with the actor selecting actions and the critic evaluating them to guide policy updates.The stated theorem establishes a stationary-point property for the resulting policy.Source: 1) Stage I–Design and Learning of… · pp. 8–9

2) Stage II–Design and Learning of Fine-Grained Level Matching Policy Within Matched Task-Resource Sets:

The fine-grained policy allocates subtasks to collaborators within sets matched by the coarse-grained policy. Its trust-informed state, restricted action space, and fairness-adjusted rewards support long-term matching.

C. Computational Complexity Analysis

The hierarchical design confines computation to coarsened graphs and matched task-resource sets, reducing matching and graph-processing burden. Dynamic coarsening adds only a lower-order cost under the stated size conditions.

  • Dynamic coarsening incurs O(|SM|X|F|I + |N|Y|F|I), which is a lower-order cost because X ≪ |SM| and Y ≪ |N|.|F| is the feature dimension and I is the number of K-means iterations.Source: C. Computational Complexity Analysis · p. 9
  • Fine-grained computation is restricted to matched task-resource sets, with the number of edges within those sets denoted by |Ep∈Cx→n∈Cy|.Source: C. Computational Complexity Analysis · p. 10
  • The hierarchical design limits coarse-grained computation to the coarsened graph and fine-grained optimization to matched task-resource sets, reducing overall matching and graph-processing burden.Source: C. Computational Complexity Analysis · p. 10

IV. PERFORMANCE EVALUATION

The evaluation is organized around convergence, robustness across problem settings, and resilience to attacks. Simulations use randomly distributed UAVs and UGVs, varied task graphs, and specified attack and learning configurations.

IV. PERFORMANCE EVALUATION

  • The evaluation asks how quickly the strategy converges, how robust it is across problem sizes and settings, and how resilient it is to attack scenarios.Source: §IV · p. 10

A. Simulation Setup

1) Parameter Settings of Tasks and Resources:

  • The task settings use Bull, Star, and Tadpole graphs, completion goals of {0.2, 0.5, 0.8}, tasks generated every 20 time slots, and resource-edge weights that differ by problem size.New tasks last 0.05 s; resource-edge weights are [0.05, 0.06] in small and [0.01, 0.02] in large problem sizes.Source: 1) Parameter Settings of Tasks and… · p. 10

2) Parameter Settings of Attacks:

  • The attack settings include ground Eves and malicious collaborators at 10% and 20% of UGVs, respectively, and On-Off attackers switching behavior in 20-time-slot cycles.The HTST-GNN uses two spatial aggregation layers; the actor and critic use 250 hidden units each, with learning rates 0.001 and 0.01, respectively.Source: 2) Parameter Settings of Attacks · p. 10

B. Comparison Strategies

The proposed hierarchical strategy combines spatial-temporal trust prediction with fairness-aware DRL; simulations report improved completion performance and robustness across tested problem sizes, fairness objectives, and attack conditions.

B. Comparison Strategies

  • The comparison baselines isolate components: GDTrust uses weighted trust, TrustGNN omits temporal aggregation, SARSA omits trust and fairness, and FairPG uses Monte Carlo policy gradients.The proposed strategy is compared with these representative dynamic collaborator-selection strategies.Source: B. Comparison Strategies · p. 10

C. Convergence Analysis (Q1)

  • The proposed strategy converges in approximately 10,000 steps and stabilizes around −0.2 for proportional fairness utility.TrustGNN and GDTrust converge more slowly; SARSA performs poorly without fairness consideration, and FairPG converges more slowly.Source: C. Convergence Analysis (Q1) · p. 10, Figure 4 · p. 11
  • For α = 2, the proposed strategy converges fastest and most stably; Bull tasks converge faster than Star and Tadpole tasks, whose utilities gradually converge.Star and Tadpole tasks initially show lower or more fluctuating utility; Bull tasks have higher connectivity.Source: C. Convergence Analysis (Q1) · pp. 10–11, Figure 4 · p. 11

D. Robustness and Parameter Sensitivity Analysis (Q2)

  • The proposed strategy achieves the highest fairness utility across small and large problem sizes under both fairness objectives, demonstrating scalability and robustness.FairPG optimizes fairness explicitly but its Monte Carlo-based policy gradient converges more slowly.Source: Figure 5 · p. 11, D. Robustness and Parameter… · p. 11
  • Thresholds around 0.4–0.6 perform strongly; ν = 0.5 balances prediction accuracy and computational efficiency.Smaller thresholds retain more unrelated graph information and increase training overhead; larger thresholds prune relevant task relationships and reduce utility.Source: D. Robustness and Parameter… · p. 11, Figure 6 · p. 12

E. Attack Resistance Analysis (Q3)

  • At 50% on-off attackers, resistance remains approximately 93%; with 3–12 Eves, secrecy rates fall, yet the proposed strategy retains stronger resilience than baselines.On-off attacks are more disruptive because intermittent malicious behavior reduces detectability and helps attackers maintain deceptive trust values.Source: Figure 7 · p. 12, E. Attack Resistance Analysis (Q3) · p. 12

V. CONCLUSION

  • 22% improvement in concurrent task completion performance over representative baselines, with approximately 93% average attack resistance even when 50% of UGVs launch on-off attacks.The conclusion reports both outcomes for the proposed strategy.Source: §V · p. 12
Loading 2610.33295v1…