Efficient routing protocols are required in opportunistic networks due to the inherent challenges of sporadic connectivity, transmission delays, and dynamic network topologies. In these networks, information is forwarded and disseminated among smart devices based on opportunistic contacts driven primarily by network dynamics and user mobility. This paper introducesLearnRouter, a dynamic reinforcement learning-based routing algorithm designed to improve the message delivery ratio while minimizing end-to-end delay in opportunistic networks. Existing routing protocols use static strategies or fixed replication schemes, leading to resource utilization and high message drop rates. LearnRouter addresses these challenges by incorporating the multi-armed bandit framework and utilizing the Upper Confidence Bound algorithm to make more intelligent and data-driven forwarding decisions. The proposed approach enables the protocol to adapt continuously to network changes while balancing the trade-off between exploration and exploitation. The simulation results show that LearnRouter surpasses end-to-end delay and message delivery ratio in comparison with Direct Delivery, Spray and Wait, CBR, SimRouter, RL-Prophet, and K-DQLR in opportunistic and dynamic environments.