Exploring Reinforcement Learning to Aid Tor Latency Performance

This paper explores the novel application of reinforcement learning to enhance Tor’s latency performance. We utilize on-the-fly training model that employs the Q-learning algorithm. The model begins with the relay selection probabilities from the Tor consensus file and progressively trains to improve its ability to construct low-latency circuits. The model’s probabilistic policy is dynamically updated based on real-time consensus file updates, aiming to adapt to the ever-changing Tor network.