Adaptive Probabilistic RREQ Rebroadcasting Using Thompson Sampling for Mobile Ad Hoc Sensor Networks

Route discovery in ad hoc on-demand distance vector (AODV)-based mobile ad hoc sensor networks relies on route request (RREQ) dissemination, which improves reachability, but generates redundant rebroadcasts, channel contention, delay, and energy waste. Fixed probabilistic rebroadcasting mitigates broadcast storms, but its performance depends on a manually selected forwarding probability applied uniformly across different local redundancy conditions. This work proposes the Thompson-sampling probabilistic AODV (TSP-AODV), a lightweight adaptive extension, in which each intermediate node selects among a small set of forwarding probability arms using a local duplicate-pressure context and delayed route reply (RREP) feedback. The reward function uses the local observation of a corresponding RREP as delayed feedback while penalising high forwarding probability in locally redundant contexts. TSP-AODV requires no additional control packets, no topology exchange, and only a small number of local Beta belief distribution parameters per node. Evaluated against AODV, fixed probabilistic rebroadcasting, counter-based suppression, and dynamic probabilistic-counter suppression over 1600 simulation runs spanning four node densities and four mobility levels, TSP-AODV achieves the lowest normalised routing overhead and the highest RREQ suppression ratio, while no statistically significant PDR difference relative to the fixed probabilistic baseline was observed under the tested conditions. End-to-end delay is also reduced significantly relative to fixed probabilistic rebroadcasting. The learned belief behaviour confirms context-dependent adaptation, with the dominant forwarding-probability arm decreasing as duplicate pressure increases. An additional 800-run sensitivity analysis characterises the PDR–NRO–delay trade-off across the tested penalty coefficients and feedback-window durations in two representative scenarios. These results are limited to the evaluated parameter grid and do not establish scenario-independent parameter robustness.