Intelligent Offloading in Vehicular Edge Computing: A Comprehensive Review of Deep Reinforcement Learning Approaches and Architectures

This academic survey examines deep reinforcement learning (DRL) approaches for optimizing computational offloading in vehicular edge computing systems. The research classifies existing DRL strategies across learning paradigms, system architectures, and optimization objectives while identifying challenges in scalability and coordination for next-generation intelligent transportation systems.

This comprehensive review addresses a critical technical challenge in modern transportation infrastructure: deciding when and where vehicles should offload computational tasks to optimize system performance. As connected and autonomous vehicles generate increasingly complex data processing requirements, centralized cloud computing becomes impractical due to latency constraints and bandwidth limitations. Edge computing—processing data closer to the source through roadside servers, peer vehicles, and UAVs—offers a solution, but requires intelligent decision-making under highly dynamic conditions.

The emergence of DRL as an offloading strategy reflects broader industry trends toward adaptive, self-optimizing systems. Traditional rule-based approaches cannot handle the heterogeneity and unpredictability of vehicular networks where vehicle density, connectivity quality, and server availability fluctuate rapidly. DRL algorithms learn optimal policies through interaction with their environment, enabling systems to improve performance over time without explicit programming for every scenario.

For developers and infrastructure providers, this research signals growing importance of distributed intelligence in transportation networks. Multi-agent DRL approaches promise better scalability than centralized solutions, while hierarchical architectures balance computational efficiency with coordination effectiveness. The survey's emphasis on reward design and coordination mechanisms indicates these remain open problems requiring further innovation.

Key challenges ahead include ensuring scalability across large vehicular networks, maintaining fairness among competing tasks and vehicles, and achieving convergence in non-stationary environments where network conditions constantly change. Future development will likely focus on federated learning approaches that preserve privacy while enabling collaborative intelligence, and hybrid systems combining DRL with traditional optimization techniques.

• →DRL enables adaptive offloading decisions in dynamic vehicular networks where traditional rule-based strategies fail to handle heterogeneous conditions.
• →Multi-agent and hierarchical DRL architectures show promise for distributed decision-making with improved scalability compared to centralized approaches.
• →Reward design and coordination mechanisms remain critical open challenges in developing effective vehicular edge computing systems.
• →Latency, energy consumption, and fairness emerge as competing optimization objectives requiring careful balancing in real-world deployments.
• →Federated learning and privacy-preserving collaborative intelligence represent key future research directions for next-generation ITS.