TRUST-AWARE ENVIRONMENT-SENSITIVE DEEP REINFORCEMENT LEARNING FOR ADAPTIVE CLUSTER HEAD SELECTION IN MOBILE AD HOC NETWORKS

Abstract

Mobile Ad hoc Networks (MANETs) suffer from performance degradation as node count and mobility increase, since every node must independently manage routing without fixed infrastructure. Clustering is a well-established remedy: nodes are grouped under an elected cluster head (CH) that coordinates intra- and inter-cluster communication, which reduces routing-table size and control overhead. Most existing CH-selection schemes score candidate nodes using a fixed combination of residual energy, node degree, mobility and distance, but they treat trust as a secondary or static attribute and re-evaluate it too infrequently to catch adaptive insider attacks such as wormhole or grey-hole routing. This paper extends a reward-optimized Deep Q-Learning (RoDQL) clustering formulation by folding a continuously-updated, feedback-driven trust score into the state representation and reward function of the learning agent, alongside energy, mobility, node-degree and distance. A guard-node subsystem audits packet-forwarding behaviour, updates each node's trust value on every routing cycle, and feeds this into the agent so that a node's suitability as CH is re-assessed as its behaviour changes, not only at cluster-formation time. The resulting framework, referred to here as Trust-Augmented RoDQL (TA-RoDQL), is presented together with its Markov Decision Process formulation, state/action/reward design, and a two-stage algorithm for cluster-head election and trust-triggered re-election. The expected benefits — more balanced cluster sizes, lower energy depletion at CHs, and faster detection and isolation of malicious nodes — are discussed against the underlying reward-shaping mechanics, and a simulation protocol is laid out so the approach can be validated empirically against classical Q-learning, DQN-only and static-weight clustering baselines.