Stochastic Reset Pathfinding: Path-Level Regret for Cascading Bandits over Graph Paths

Short summary

Stochastic Reset Pathfinding (SRP) formalizes episodic learning on directed graphs with unknown edge success probabilities, where any edge failure resets the agent to the source. The authors prove the optimal policy is open-loop, placing SRP within combinatorial cascading bandits, and propose PathUCB and PathTS algorithms with a novel path-level regret bound. Experiments across quantum-network, grid-world, and Erdos-Renyi domains show PathTS performs best empirically, though adversarial instances can prevent convergence.

• •SRP models pathfinding on graphs with unknown edge failure probabilities and global resets
• •PathUCB achieves a path-level regret bound decomposing over suboptimal paths via per-path complexity
• •PathTS is recommended as practical default but can fail to converge on adversarial instances

Generated with AI, which can make mistakes.