PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Multi-Discriminator WGAN Framework for Resource-Constrained Internet of Vehicles Using Reinforcement Learning and Game Theory

Farhoud Jafari Kaleibar, Amr M. Zaki, Marin Litoiu

arXiv:2610.10926v1cs.LGcs.DC

TL;DR

IoV GAN deployment requires resource orchestration as mobility and connectivity change, while traditional architectures face competing accuracy, resource, latency, and overhead objectives. The paper proposes MD-WGAN, coordinating discriminator selection and training epochs through reinforcement learning and game theory. Its conclusion reports improved prediction accuracy, reduced resource consumption, and latency optimization.

  • Problem

    Traditional GAN architectures in IoV face mobility, resource, and latency constraints while the objectives of accuracy, resource use, delay, and communication overhead compete.

    Source: §I · p. 1

  • Method

    MD-WGAN combines reinforcement-learning discriminator selection with game-theoretic epoch allocation in a unified optimization framework for vehicular constraints.

    Source: §I · p. 2

  • Results

    Experiments report improved prediction accuracy, reduced resource consumption, and consistent latency optimization with the proposed framework.

    Source: §VI · p. 14

  • Takeaways & Limitations

    The paper presents joint optimization of adversarial learning and deployment constraints as its approach to IoV network management.

    Source: §I · p. 2

Abstract

from arXiv · show

Managing machine learning workloads as a network service introduces a resource-orchestration problem distinct from conventional model training; which nodes should be allocated to a task, how communication and computation budgets should be divided among them, and how service quality should be sustained as connectivity and node availability change with mobility. Deploying Generative Adversarial Networks (GANs) in Internet of Vehicles (IoV) environments is a demanding instance of this problem; resource constraints, dynamic network topologies, and competing optimization objectives mean that traditional GAN architectures cannot simultaneously achieve high accuracy, efficient resource use, low delay, and low communication overhead. This paper introduces an adaptive multi-discriminator Wasserstein GAN (MD-WGAN) framework that integrates reinforcement learning with game-theoretic coordination to address these challenges jointly. In our framework, roadside units host generators paired with Deep Q-Network (DQN) agents that select discriminator subsets and manage distributed training across mobile vehicular nodes, while a game-theoretic coordination step allocates training epochs between generators and discriminators. A unified optimization objective ties adversarial learning quality to resource efficiency, communication overhead, and latency under vehicular constraints, allowing the framework to continuously adapt its training behavior as network conditions change. Evaluation on real-world NGSIM trajectory data shows that the framework attains prediction accuracy comparable to state-of-the-art GAN baselines - the lowest RMSE (1.029) and MAE (0.894) among all evaluated methods - while markedly improving resource efficiency: average CPU utilization is reduced by roughly 28% and mean memory usage by roughly 6%, at competitive communication overhead and latency.

I. INTRODUCTION

The introduction frames IoV GAN deployment as a resource-orchestration problem under mobility, intermittent connectivity, and competing objectives. It introduces MD-WGAN, combining multi-agent reinforcement learning and game-theoretic coordination to jointly address these constraints.

1) Generative Adversarial Networks:

  • GANs can augment scarce vehicular data and synthesize rare scenarios, while adversarial training pits a data-generating network against one that distinguishes synthetic from real samples.Synthetic samples may also be shared in place of raw sensor data, addressing privacy concerns.Source: §I · p. 1, 1) Generative Adversarial Networks · p. 2

I. INTRODUCTION

  • The framework targets the joint optimization of latency, accuracy, resource efficiency, and network overhead in dynamic vehicular networks.It maps generators to roadside units and discriminators to vehicles, with reinforcement-learning selection and game-theoretic epoch allocation.Source: §I · pp. 1–2

2) Wasserstein GAN:

  • WGAN replaces Jensen–Shannon divergence with Wasserstein distance and uses a 1-Lipschitz critic, providing smoother gradients and more stable convergence.The paper enforces the Lipschitz constraint with a gradient penalty and calls WGAN critics “discriminators” for consistency with IoV terminology.Source: 2) Wasserstein GAN · p. 2

3) Internet of Vehicles Characteristics:

  • IoV training must accommodate intermittent vehicle connections, heterogeneous computing and communication resources, and fluctuating network conditions without always-on centralized infrastructure.These constraints motivate a distributed, resource-aware, multi-discriminator design.Source: 3) Internet of Vehicles Characteristics · p. 2

B. Related Work

  • The related-work review spans GAN-based vehicular networks, generative AI beyond GANs, and federated learning applications.It positions the paper against existing approaches that do not jointly address learning efficiency, overhead, latency, and resource use.Source: §I · p. 1, B. Related Work · p. 2

1) GAN-based Vehicular Networks:

  • HVS-GAN focuses on scheduling, while Liao et al. face computational complexity; GIDS omits resource efficiency, and Ioannou and Vassiliou evaluate accuracy alongside energy cost.The paper identifies multi-objective optimization as a remaining need for resource-constrained vehicular settings.Source: 1) GAN-based Vehicular Networks · pp. 2–3

2) Generative AI based Vehicular Networks:

The section reviews generative-AI and federated-learning approaches before describing the MD-WGAN system model. Its architecture pairs roadside-unit generators with vehicle discriminators over V2I links, using trajectory contexts to generate next-step features.

2) Generative AI based Vehicular Networks:

  • DiffusionDrive reduces denoising steps 10× but still requires substantial inference resources; other generative approaches also lack joint optimization of prediction accuracy, computation, and communication.GAI-IoV focuses on content generation, while conditional VAEs are more computationally efficient than some alternatives.Source: 2) Generative AI based Vehicular… · p. 3

3) Federated Learning in IoV:

  • Prior federated-learning approaches address client selection, traffic management, or mobility, but are described as limited by learning trade-offs, scalability, overhead, or centralized aggregation bottlenecks.The proposed framework targets learning efficiency, resource use, network overhead, and latency together.Source: 3) Federated Learning in IoV · p. 3

III. SYSTEM MODEL

  • Each roadside unit hosts a generator and local controller, while vehicles act as discriminators evaluating generated data against real traffic samples over V2I links.The controller’s coordination mechanisms are detailed in Section IV.Source: §III · p. 3

1) IoV Network Architecture:

2) WGAN Training Model:

  • A generator maps past trajectory context and latent noise to a predicted next-step feature vector, while discriminators evaluate real and generated samples under the WGAN objective.Training pairs use a context and its ground-truth continuation; the section also identifies a gradient penalty and distributions used in the objective.Source: 2) WGAN Training Model · p. 4

B. Communication Model

The communication and computation model characterizes wireless links, training-cycle transmission costs, and node resource constraints. It distinguishes training synchronization latency from inference-time latency.

1) Wireless Channel Model:

  • The wireless channel model uses log-distance path loss, with gain depending on generator–discriminator distance, path-loss exponent, and reference gain.The paper identifies these as d_i,j, β, and h0, respectively.Source: 1) Wireless Channel Model · p. 4

2) Data Transmission:

  • Data transmission rate is modeled with Shannon–Hartley channel capacity using bandwidth, transmission power, and noise power spectral density.The communication model also defines latency and network traffic overhead.Source: 2) Data Transmission · p. 4
  • Communication latency measures a federated training and synchronization round trip, not downstream generator inference time.The round trip covers sample broadcast, discriminator evaluation, feedback aggregation, and generator update.Source: 2) Data Transmission · p. 4

C. Computation Model; 2) Resource Utilization:

IV. ADAPTIVE MULTI-DISCRIMINATOR FRAMEWORK

The framework links adaptive decision-making, game-theoretic coordination, distributed training, and monitoring in a recurring cycle. Its joint objective minimizes Wasserstein distance subject to IoV resource and network constraints.

IV. ADAPTIVE MULTI-DISCRIMINATOR FRAMEWORK

  • The framework is presented as a recurring coordination workflow that integrates reinforcement learning for local optimization with game theory for global resource management.Figure 2 depicts the proposed approach’s workflow.Source: A. Overview: The Adaptation Cycle · p. 5, Figure 2 · p. 6

A. Overview: The Adaptation Cycle

  • The four-step adaptation cycle selects discriminators and tunes parameters, allocates epochs, performs distributed training and synchronization, then feeds monitoring metrics into the next round.The loop repeats until convergence criteria are satisfied.Source: A. Overview: The Adaptation Cycle · p. 5

B. Joint Optimization Objective

  • The joint optimization minimizes Wasserstein distance while accounting for resource and network constraints for each generator and its associated discriminator subset.The objective’s five coefficients form a convex combination, so increasing one objective’s weight reduces the relative influence of the others.Source: B. Joint Optimization Objective · p. 5

1) Objective Function Components: •

The objective balances adversarial learning against resource and network costs, while the adaptive decision engine selects participating discriminators using feedback. Its decision variables include model parameters, discriminator selection, and resource allocation.

1) Objective Function Components: •

  • The objective combines WGAN training loss with normalized CPU, memory, communication-delay, and transmission-volume costs.The delay term represents average normalized communication delay; lower values indicate faster feedback exchange.Source: 1) Objective Function Components: • · p. 6

2) Decision Variables:

  • The optimization decides model parameters, each generator’s discriminator subset, and computing resources allocated to each discriminator.These decisions jointly determine which nodes participate and how resources are assigned.Source: 2) Decision Variables · p. 6

C. Step 1: Adaptive Decision Engine; 1) RL Formulation: •

  • Each generator’s DQN agent learns discriminator subset selection from environmental feedback, using a seven-dimensional state and per-discriminator Q-values to select top-k candidates.The state includes loss, score, latency, CPU, memory, traffic, and a prior-round discriminator score; the approach avoids enumerating subsets but does not model interactions among simultaneously selected discriminators.Source: C. Step 1: Adaptive Decision Engine · p. 6, 1) RL Formulation: • · p. 6

1) RL Formulation: •

  • The DQN uses fully connected layers with ReLU activations and experience replay, selecting actions greedily during exploitation.The stated layer widths are [128, 128, |A|], and the replay buffer has capacity 10,000.Source: 1) RL Formulation: • · pp. 6–7

2) Bootstrapping and Discriminator Selection:

Discriminator selection begins with exploration and performance-based ranking, while a per-round reward balances training improvement against resource and communication costs. The selected subset is the decision output; other training configurations remain fixed.

2) Bootstrapping and Discriminator Selection:

  • Initialization explores available discriminators, ranks them by convergence speed and loss reduction, then selects the top-k for active training.The initial ε-greedy exploration rate is 1.0.Source: 2) Bootstrapping and Discriminator… · p. 7

3) Reward Formulation:

  • The per-round reward balances GAN-loss improvement against normalized latency, network traffic, CPU, and memory usage.The reward weights sum to one, and the agent maximizes expected discounted reward.Source: 3) Reward Formulation · p. 7

4) Exploration Schedule:

  • Exploration decreases during training but retains a minimum rate to remain responsive to network dynamics.The exploration schedule values are specified in Table III.Source: 4) Exploration Schedule · p. 7

5) Decision Output:

  • The decision engine outputs the discriminator subset for distributed training, while learning rate, batch size, and synchronization interval are fixed.Joint per-round selection of those training configurations is left for future work.Source: 5) Decision Output · p. 7

D. Step 2: Game-Theoretic Best-Response Coordination

The coordination step treats generator and discriminator epoch allocation as a strategic game, balancing training effectiveness against resource costs and interference. A bounded best-response search selects an allocation that improves coordination without claiming convergence to a Nash equilibrium.

D. Step 2: Game-Theoretic Best-Response Coordination

1) Utility Functions:

  • Utilities balance epoch-related training effectiveness against computational or communication costs and interference from the opposing player's actions.The effectiveness terms have diminishing returns, resource costs scale linearly with epochs, and interference is proportional to the opponent's effectiveness.Source: 1) Utility Functions · p. 7, 2) Component Instantiation · p. 7

2) Component Instantiation:

  • The effectiveness functions model progressively smaller convergence gains from additional discriminator or generator epochs.The resource-cost terms are linear in epoch count and scaled relative to effectiveness.Source: 2) Component Instantiation · p. 7

3) Best-Response Search Procedure:

  • The bounded search selects the candidate pair with the higher joint payoff, rather than iterating until a fixed point.The authors therefore describe it as best-response coordination, not a converged Nash equilibrium.Source: 3) Best-Response Search Procedure · p. 7

4) Effect on Coordination:

  • The coordination step provides a principled epoch distribution instead of a fixed schedule and reduces conflicting parameter updates from independent training.The passage also characterizes the step as improving resource allocation across nodes.Source: 4) Effect on Coordination · p. 7

E. Step 3: Distributed Training and Synchronization

Distributed WGAN training uses a dynamically selected discriminator subset and allocated epochs, with generators receiving ensemble feedback and discriminators trained independently. Monitoring of network, resource, and training metrics feeds the next adaptation cycle; the evaluation setup combines NS3 simulation with a PyTorch training pipeline.

E. Step 3: Distributed Training and Synchronization

2) Discriminator Training:

  • Each discriminator is trained independently using the Wasserstein loss with gradient penalty.The passage defines the real sample as drawn from the data domain and λGP as the WGAN-GP coefficient.Source: 2) Discriminator Training · p. 8

F. Step 4: Monitoring and Convergence Criteria

  • Monitoring tracks latency, bandwidth, topology and connection stability, CPU and memory use, and generator and discriminator training quality.Both generator and discriminator losses must stabilize before the training cycle terminates, and monitoring feeds back into the state vector.Source: F. Step 4: Monitoring and Convergence… · p. 8

A. Experiment Setup

  • The evaluation simulation integrates NS3's 802.11p network stack with a PyTorch adaptive GAN pipeline and SUMO-generated urban vehicle mobility.Each curve averages five independent runs with different random seeds.Source: A. Experiment Setup · p. 8
  • Unless otherwise noted, experiments use equal weights of 0.2 for each of the five reward components.RL-agent and game-theoretic coordination hyperparameters are reported in Tables III–IV.Source: A. Experiment Setup · p. 8, Table III · p. 8

B. Dataset

The study uses enhanced NGSIM trajectory data organized into non-overlapping context–target continuations, and compares GAN-based methods on held-out prediction accuracy and system-resource metrics. Its baselines span adaptive-control, federated, and communication-efficient approaches, while the comparison is limited to GAN-based methods.

B. Dataset

  • Enhanced NGSIM samples comprise five trajectory features organized as past-context windows and subsequent-vector targets, with sequence-disjoint training and held-out test sets.The data are preprocessed for multi-discriminator GAN training under non-IID conditions.Source: B. Dataset · pp. 8–9

C. Evaluation Metrics

  • RMSE and MAE assess prediction deviation on held-out continuations, while latency and network overhead measure communication performance.CPU use and memory consumption track resource use across participating nodes.Source: C. Evaluation Metrics · p. 9

D. Baseline Comparisons

  • The comparison includes random parameter selection, HVS-GAN, and Early Update MD-GAN, and is restricted to GAN-based baselines.The authors state that the study targets the resource-aware coordination layer within GAN training rather than comparisons across generative-model paradigms.Source: D. Baseline Comparisons · p. 9

F. Experimental Results

Across the evaluated methods, adaptive MD-WGAN achieves the lowest reported RMSE and MAE, while the results also describe its training-cost progression, resource use, network overhead, and latency. The reported comparisons show different trade-offs across those system measures.

1) Multi-Objective Performance Evaluation:

  • Training cost shifts from volatile exploration before roughly round 12 or 13 to a stable, near-optimal phase after roughly round 57.The learning phase is approximately rounds 13–57, with improving cost stability and a consistent downward trend.Source: Figure 3 · p. 10, 1) Multi-Objective Performance… · p. 10

2) Accuracy Performance Analysis:

3) Resource Utilization:

  • During active training, adaptive MD-WGAN memory use is about 8,000–12,000 MB, and its CPU use is 500–1,400%.Early Update MD-GAN memory peaks at approximately 14,500 MB; its CPU and HVS-GAN CPU frequently spike to 1,500–2,200%.Source: 3) Resource Utilization · pp. 10–11, Figure 5 · p. 11

4) Network Traffic Overhead:

  • Network overhead reaches approximately 1.2 million bytes for adaptive MD-WGAN versus approximately 0.9 million bytes for Early Update MD-GAN after 100 seconds.The proposed method is described as the second-best performer, while HVS-GAN reaches nearly 4.8 million bytes.Source: Figure 6 · p. 11, 4) Network Traffic Overhead · p. 11

5) Latency:

  • Adaptive MD-WGAN has a median latency of approximately 9.9 seconds, and its training-iteration chart approaches 7 seconds after iteration 50.HVS-GAN's median is approximately 9.7 seconds; the proposed approach is described as having the tightest latency distribution.Source: 5) Latency · p. 11, Figure 7 · p. 12

6) Scalability Analysis:

The scalability analysis characterizes training and communication costs alongside adaptive decision-making overhead. It reports bounded coordination costs as the network grows, while communication traffic scales with generator nodes, selected discriminators, parameter size, and synchronization steps.

6) Scalability Analysis:

  • With a bounded local discriminator pool, per-decision-cycle DQN complexity remains O(1) as generator and discriminator counts grow.Game-theoretic best-response evaluation is also constant with network scale because its strategy sets are bounded.Source: 6) Scalability Analysis · p. 12
  • The analysis identifies gradient calculation and backpropagation for generators and discriminators as the most computationally intensive training tasks.The section frames training complexity as dependent on node count, discriminator count, and data size.Source: Computational Complexity (Ttrain) · p. 12, 6) Scalability Analysis · p. 12
  • Discriminator optimization scales with model size and batch size, while parallel execution makes its runtime independent of the number of nodes.Generator loss and gradient updates also scale with model parameters and batch size.Source: Computational Complexity (Ttrain) · p. 12
  • Per synchronization step, communication complexity scales with generator nodes, selected discriminators, and model parameters; total traffic also scales with synchronization steps.The analysis sums generator-parameter synchronization and discriminator-gradient exchange.Source: Communication Complexity (Ccomm) · pp. 12–13

7) Application-Specific Multi-Objective Performance Validation:

The application-specific evaluation varies objective weights to prioritize different IoV needs, and the conclusion reports improved prediction accuracy, resource consumption, and latency relative to state-of-the-art approaches. The paper identifies compression, emerging 5G/6G support, and specialized vehicular adaptations as future directions.

7) Application-Specific Multi-Objective Performance Validation:

  • Across four application scenarios, the cost function reweights latency, GAN quality, CPU and memory use, or network traffic; Figure 10 compares metrics across apps.The reported results validate adaptation to task requirements.Source: 7) Application-Specific… · p. 13, Figure 10 · p. 14

VI. CONCLUSION

  • The conclusion reports improved prediction accuracy, reduced resource consumption, and consistent latency optimization compared with state-of-the-art approaches.It presents the framework as combining adversarial learning with reinforcement learning and game-theoretic coordination.Source: §VI · p. 14
  • Future work includes advanced compression, support for emerging 5G/6G standards, and adaptations for autonomous-vehicle coordination and smart traffic management.Source: §VI · p. 14
Loading 2610.10926v1…