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

The gist Managing machine learning workloads as a network service introduces a resource-orchestration problem distinct from conventional model training, which this framework addresses by integrating reinforcement learning with game theory to optimize performance under vehicular constraints.

Framework Overview

The proposed adaptive multi-discriminator Wasserstein GAN (MD-WGAN) framework integrates reinforcement learning with game-theoretic coordination to address challenges like high accuracy, efficient resource use, low delay, and low communication overhead in Internet of Vehicles (IoV) environments <ref:2610.10926#pg2>. In this framework, roadside units host generators paired with Deep Q-Network (DQN) agents that select discriminator subsets and manage distributed training across mobile vehicular nodes <ref:2610.10926#pg4>. A game-theoretic coordination step allocates training epochs between generators and discriminators to improve convergence under network dynamics <ref:2610.10926#pg4>. The framework employs a four-step cyclical methodology that continuously adapts to changing network conditions <ref:2610.10926#pg6>.

System Model and Constraints

The system architecture maps the generator and discriminator roles onto the physical IoV topology, where stationary roadside units (RSUs) act as generators and mobile vehicles act as discriminators evaluating generated data under mobility constraints <ref:2610.10926#pg4>. The training objective is formulated to minimize the Wasserstein distance while satisfying resource and network constraints in the IoV environment <ref:2610.10926#pg6>. The joint optimization problem minimizes L(t)W GAN while subject to constraints on CPU usage, memory usage, communication latency, and network traffic overhead <ref:2610.10926#pg6>.

Adaptive Decision Engine (Reinforcement Learning)

The Adaptive Decision Engine is the intelligent core of the framework, where each generator node maintains an independent Deep Q-Network (DQN) agent that learns a policy for discriminator subset selection from environmental feedback <ref:2610.10926#pg8>. The state space st aggregates metrics such as average discriminator loss, communication latency, CPU utilization, memory consumption, network traffic, and the discriminator-side average score from the previous round <ref:2610.10926#pg8>. The action space A is defined by the number of discriminators in the local pool assigned to generator gi (Klocal), and the DQN outputs a Q-value per candidate discriminator, selecting top-k discriminators as active subset D(t)gi <ref:2610.10926#pg8>. The reward formulation balances training effectiveness against resource and communication cost, defined by Eq. (16) <ref:2610.10926#pg8>.

Game-Theoretic Coordination

The framework models the interaction between discriminators and generators as a strategic game to allocate training epochs, solved via a bounded best-response search over discretized strategy sets <ref:2610.10926#pg9>. The utilities UD(ED, EG) and UG(ED, EG) balance training effectiveness against resource cost and interference terms reflecting the opponent’s actions <ref:2610.10926#pg9>. The best-response coordination step selects the pair (E∗D, E∗G) with the higher joint payoff UD + UG for that round <ref:2610.10926#pg9>. This coordination step improves resource allocation and provides a principled epoch distribution between generator and discriminator in place of a fixed schedule <ref:2610.10926#pg9>.

Performance Evaluation

Experimental validation demonstrates that the proposed adaptive MDWGAN framework achieves the lowest RMSE (1.029) and MAE (0.894) among all evaluated methods while markedly improving resource efficiency, such as reducing average CPU utilization by roughly 28% and mean memory usage by roughly 6% <ref:2610.10926#pg2>. The framework exhibits better convergence performance compared to state-of-the-art approaches, maintaining consistent improvement throughout training <ref:2610.10926#pg8>. Furthermore, the framework shows superior latency reduction capability as training progresses over time <ref:2610.10926#pg8>. The scalability analysis indicates that the computational complexity remains bounded as the network scales because the per-decision-cycle DQN complexity is O(Klocal) = O(1) with respect to NG and K <ref:2610.10926#pg14>.

Application Scenarios

The framework's performance can be adapted to specific task requirements by adjusting the weighting coefficients (αG, α1, α2, α3, α4) in the cost function for four distinct scenarios <ref:2610.10926#pg8>. These scenarios include a Delay-Sensitive Application prioritizing minimizing communication latency with a high value for alpha3 = 0.4 <ref:2610.10926#pg8>, an Accuracy-Oriented Application maximizing GAN’s generation quality with alphaG = 0.4 <ref:2610.10926#pg8>, a Resource-Constrained Application focusing on minimizing consumption of limited computational resources with alpha1 and alpha2 set to 0.3 <ref:2610.10926#pg8>, and a Network Overhead-Oriented Application aiming to minimize total network traffic with alpha4 = 0.4 <ref:2610.10926#pg8>. The results validate the cost function’s ability to adapt the framework’s behavior to specific task requirements <ref:2610.10926#pg8>.

The demonstrated effectiveness establishes a foundation for next-generation AI-driven vehicular network optimization, contributing to intelligent transportation systems advancement through improved efficiency, reliability, and realtime performance. The framework provides a practical solution for vehicular network operators and ITS developers seeking to leverage machine learning capabilities while maintaining operational efficiency, making it particularly valuable for connected and autonomous vehicles, smart city infrastructure, and cooperative transportation systems.

--- Page 14 ---

<ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg4> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.10926#pg8> <ref:2610.10926#pg9> <ref:2610.10926#pg6> <ref:2610.10926#pg8> <ref:2610.

1. textbfAdaptive Multi-Discriminator WGAN (MD-WGAN) Framework for Joint Optimization: A unified optimization objective ties adversarial learning quality to resource efficiency, communication overhead, and latency under vehicular constraints. This allows the framework to continuously adapt its training behavior as network conditions change by jointly optimizing the Wasserstein distance with resource and network cost terms in Equation (8).

2. textbfHierarchical Coordination Protocol: A hierarchical coordination protocol combining Deep QNetwork (DQN) agents for discriminator subset selection with game-theoretic best-response coordination for training epoch allocation. This enables the system to adapt to network conditions, resource availability, and data distribution shifts by using DQN for local RL decisions and best-response search for global epoch allocation.

3. textbfReal-Time Resource Allocation: The reinforcement learning agent learns a policy for discriminator subset selection from environmental feedback, as defined in Step 1 of the adaptation cycle. This allows the system to dynamically select discriminators based on observed metrics like average discriminator loss, average discriminator score, communication latency, CPU utilization, memory consumption, network traffic (State space vector st in Eq. 14).

4. textbfResource-Aware Training: The objective function minimizes the Wasserstein distance while explicitly penalizing resource usage through terms like CPU Utilization (α1 · C(t) Dgi Cmax) and Memory Consumption (α2 · M(t) Dgi Mmax). This ensures that training continues while adhering to physical constraints, preventing resource exhaustion.

5. textbfLatency-Driven Optimization: The Q-learning cost function includes a specific parameter designed to penalize high latency, as shown in Figure 7's progression chart where the proposed approach demonstrates a distinctive downward trend in latency as training progresses. This provides an intelligent, active optimization mechanism that is essential for real-world, dynamic IoV environments.

6. textbfScalable Distributed Training: The communication complexity analysis shows that the total communication traffic is Ctotal = Ssync · Ccomm = O(Ssync · NG · KS · d). This demonstrates how the framework manages computational and communication costs as the system scales, provided Intelligent Communication Scheduling minimizes synchronization steps (Ssync) and active discriminators (K).

7. textbfApplication-Specific Prioritization: The framework can be configured for specific tasks by adjusting cost function coefficients (e.g., setting α3 to 0.4 for a Delay-Sensitive Application). This allows the system to prioritize minimizing communication latency, maximizing GAN’s generation quality, or focusing on resource minimization based on the operational needs of the vehicular application.

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.