Deep RL with Optimization-Based Refinement for Mobility Prediction in Smart Cities Using MEC and V2X Communication

Abstract

The rapid growth of smart cities and the proliferation of Vehicle-to-Everything (V2X) communication networks have created an urgent need for traffic prediction systems that are simultaneously accurate, adaptive to non-stationary conditions, and deployable within the stringent latency constraints of Multi-Access Edge Computing (MEC) platforms. This thesis proposes and validates a hybrid two-stage framework for real-time directional traffic trend prediction. In the first stage, a Proximal Policy Optimization (PPO) agent equipped with a Temporal Convolution Network encoder and a multi-head self-attention spatial integration module is trained offline on a 200,000-timestep synthetic V2X dataset generated over the real urban road network of Nördlingen, Germany, learning to classify the dominant directional traffic surge (North, East, South, or West) over a 60-minute prediction horizon from 15 roadside unit observations. In the second stage, the frozen PPO policy is augmented at inference time by a Genetic Algorithm (GA) that continuously optimises a 4-dimensional logit bias vector against a sliding window of recent prediction outcomes, correcting for

systematic directional errors introduced by abrupt traffic regime shifts without any gradient- based retraining. Experimental evaluation against supervised LSTM, ARIMA, Kalman filter, and random baselines demonstrates that the PPO+GA system achieves 78.0% accuracy and 79.26% Macro-F1 on the static test set, and maintains 46.8% accuracy across four abrupt regime transitions a 50% relative improvement over the frozen LSTM (31.2%). A closed-loop validation in the SUMO traffic simulator on the Belgrade road network, with 74 signalized intersections, yields a 33.3% vehicle throughput gain and a 10.4% average speed improvement over fixed-time signal control without any retraining, demonstrating cross- network generalization. All results confirm that the proposed framework meets the computational requirements of real-time MEC deployment with an inference latency of 54.59 ms, well within the 60-second operational window.