Preventing malicious nodes in ad hoc networks using reinforcement learning

This paper proposes an enhancement to an existing reputation method for indicating and avoiding malicious hosts in wireless ad hoc networks. The proposed method combines a simple reputation scheme with a reinforcement learning technique called the on-policy Monte Carlo method where each mobile host distributedly learns a good policy for selecting neighboring nodes in a path search. Simulation results show that the reputation scheme combined with the reinforcement learning can achieve up to 89% and 29% increase in throughput over the reputation only scheme for the static and dynamic topology case, respectively.