AI Research Analysis

The framework uses a hierarchical approach: inter-task budget allocation by RSUs and intra-task rank selection by vehicles. This ensures efficient resource distribution and adaptive fine-tuning in dynamic IoV settings. This strategy enables efficient inter-task coordination and resource sharing, improving overall performance. By dynamically reallocating resources based on task demands and priorities, it ensures timely execution, energy efficiency, and high model accuracy, thereby enhancing system responsiveness and user experience.

A novel UCB-DUAL algorithm for constrained multi-armed bandit (MAB) problems enables adaptive exploration under per-task energy budgets, achieving provable sublinear regret. It combines upper confidence bound exploration with dual-variable updates for decentralized, energy-aware rank adaptation. Each agent (vehicle) independently selects its LoRA rank based on an objective function that balances reward, energy cost, and an exploration bonus, ensuring robustness in dynamic federated settings.

LoRA Rank Impact on Performance

Rank	Best Accuracy (%)	Latency (s)	Energy (J)
1	73.329	30.525	608.568
8	81.443	32.417	646.270
200	83.069	67.809	1346.458
Conclusion: Higher ranks improve accuracy but significantly increase latency and energy consumption. Adaptive rank selection is key for efficiency.

Leveraging Low-Rank Adaptation (LoRA), the system allows vehicles to adaptively select optimal ranks (η) for fine-tuning. This mechanism balances model capacity (accuracy) with computational/communication overhead (latency, energy). This adaptive rank selection is crucial for maintaining a balance between performance and cost in practical federated settings, especially given heterogeneous device capabilities and dynamic network conditions.

To address client mobility and intermittent connectivity in IoV, a fault-tolerant scheduling strategy anticipates disconnections and selects fallback actions (early upload, task migration, abandonment) to preserve training utility and minimize cost. This ensures resilient training under dynamic vehicle mobility and avoids loss of valuable gradients.