Optimizing task offloading in MIMO-enabled vehicular networks through deep reinforcement learning

Mobile Edge Computing (MEC) effectively alleviates the computational burden faced by vehicles in processing compute-intensive tasks due to resource limitations. However, traditional approaches typically employ coarse-grained task offloading strategies that utilize sequential protocols and discrete action spaces, resulting in high latency and increased energy consumption. These limitations render such strategies unsuitable for real-time applications. To address these challenges, an innovative computation offloading strategy is proposed, specifically designed to minimize the long-term average computation cost in a multi-vehicle, multi-server Internet of Vehicles (IoV) system. The MEC system model is constructed using Multiple-Input Multiple-Output (MIMO) technology, which facilitates simultaneous uplink transmissions from all vehicles, significantly reducing the time required for data uploads. Subsequently, a continuous action space is adopted to enhance both the flexibility and precision of decision-making. Additionally, Batch-Constrained Q-learning (BCQ) is introduced to further constrain the actions taken by the policy, mitigating overly optimistic estimates through a batch constraint mechanism. Finally, the Twin Delayed Deep Deterministic Policy Gradient with Batch-Constrained Q-learning (TD3BCQ) framework is developed to enable fine-grained decision-making for local execution and power allocation during task offloading within a continuous action space. Experimental results demonstrate that the proposed scheme achieves a more balanced offloading strategy and better exploits the available computing resources, leading to an approximate 20% improvement compared to the baselines.