Adaptive Path Selection in Graphs with Dynamic Action Space via Multi-Armed Bandits

The transport network (RT) in mobile networks is a critical component for serving ultra-reliable low-latency communication applications, such as cloud gaming and autonomous vehicles. It is noted in the literature the absence of solutions that simultaneously deal with the dynamic variability of routes in the face of router failures and their impacts on QoS. The work models RT as a graph and proposes an adaptive path selection approach based on Multi-Armed Bandits (MAB), with a dynamic action space and a structural adaptation mechanism in the face of failures and route recoveries. Three policies are evaluated: epsilon-greedy, Upper Confidence Bound (UCB) and Thompson Sampling (TS), in simulations with failure cycles. TS stood out widely, maintaining an average reward between 0.85 and 0.95, delay close to 10 ms, and reliability between 97% and 99%, outperforming the other methods.