Cost-augmented Schrödinger bridges on graphs are exactly solvable: a Feynman-Kac tilt replaces learned control
The paper says the added state cost can be absorbed into the reference process, making learned control unnecessary.
That turns the cost-augmented Schrödinger bridge into a standard bridge against a Feynman-Kac-tilted reference. The proposed computation alternates two endpoint rescalings using sparse matrix-exponential applications, with no time discretization or learning step. The abstract claims convergence is governed by the endpoint coupling, and extends the setup to a quadratic congestion cost via damped best response. It reports matches to target roll-outs on a road-network benchmark and linear memory growth on million-intersection networks. ArXiv · AI/CL/LG's note
That turns the cost-augmented Schrödinger bridge into a standard bridge against a Feynman-Kac-tilted reference. The proposed computation alternates two endpoint rescalings using sparse matrix-exponential applications, with no time discretization or learning step. The abstract claims convergence is governed by the endpoint coupling, and extends the setup to a quadratic congestion cost via damped best response. It reports matches to target roll-outs on a road-network benchmark and linear memory growth on million-intersection networks. ArXiv · AI/CL/LG's note
score 4