Adaptive Reinforcement Learning for Reliable and Sustainable Intelligent Vehicular Communication

Vehicular ad hoc networks (VANETs) require routing decisions that balance connectivity, link reliability and energy efficiency under continuously changing mobility and density. Conventional clustering-based routing combines these objectives using weights fixed in advance, which cannot reflect the changing relative importance of the objectives as conditions vary. This paper presents ISPY, a hybrid reinforcement learning framework for cluster-based routing in which the objective weighting is itself the learned quantity. A dueling double deep Q-network selects a discrete operating mode from a seven-dimensional macroscopic state comprising vehicle density, mean speed, speed variation, hop count, residual energy, energy efficiency and packet delivery ratio, while a Dirichlet-based continuous actor generates a normalised weight vector over anchor connectivity, link reliability and link persistence. Candidate next hops are ranked by a cost-based priority function and selected by minimisation. The reward combines packet delivery ratio, energy efficiency, and forwarding stability with coefficients 0.5, 0.3, and 0.2. The framework is evaluated on a vehicular mobility and communication trace segmented into 300 temporally aligned snapshots, compared against a fixed equal-weight policy that isolates the contribution of weight adaptation and against the conventional schemes N-HOP, VMaSC, and DMCNF under identical snapshots, transmission range, and link metric inputs. Training converges under early stopping at episode 130, with an average composite reward of 0.8161 against 0.8158 for the fixed-weight policy; this difference of 0.0003 is reported as a convergence signal. The learned weights vary systematically with network state. Comparison against learning-based baselines and a component-wise ablation are identified as necessary further work.