Abstract
Environment sensing and fusion via onboard sensors are envisioned to bewidely applied in future autonomous driving networks. This paper considers avehicular system with multiple self-driving vehicles that is assisted bymulti-access edge computing (MEC), where image data collected by the sensors isoffloaded from cellular vehicles to the MEC server usingvehicle-to-infrastructure (V2I) links. Sensory data can also be shared amongsurrounding vehicles via vehicle-to-vehicle (V2V) communication links. Toimprove spectrum utilization, the V2V links may reuse the same frequencyspectrum with V2I links, which may cause severe interference. To tackle thisissue, we leverage reconfigurable intelligent computational surfaces (RICSs) tojointly enable V2I reflective links and mitigate interference appearing at theV2V links. Considering the limitations of traditional algorithms in addressingthis problem, such as the assumption for quasi-static channel stateinformation, which restricts their ability to adapt to dynamic environmentalchanges and leads to poor performance under frequently varying channelconditions, in this paper, we formulate the problem at hand as a Markov game.Our novel formulation is applied to time-varying channels subject to multi-userinterference and introduces a collaborative learning mechanism among users. Theconsidered optimization problem is solved via a driving safety-enabledmulti-agent deep reinforcement learning (DS-MADRL) approach that capitalizes onthe RICS presence. Our extensive numerical investigations showcase that theproposed reinforcement learning approach achieves faster convergence andsignificant enhancements in both data rate and driving safety, as compared tovarious state-of-the-art benchmarks.