DataSP: A Differentiable All-to-All Shortest Path Machine Learning Algorithm to Facilitate Learning Latent Costs from Trajectories

A novel method called DataSP is proposed to learn latent costs from observed trajectories by encoding them into frequencies of observed shortcuts. DataSP is a differentiable adaptation of the Floyd-Warshall algorithm that enables informative backpropagation through shortest-path computation. It facilitates predicting likely trajectories and inferring probable destinations or future nodes.