Find in Library

We present here a model-free method for learning actions that lead to an all-source-all-destination shortest path solution. We motivate our approach in the context of biological learning for reactive control. Our method involves an agent exploring an unknown world with the objective of learning how to get from any starting state to any goal state in shortest time without having to run a path planning algorithm for each new goal selection. Using concepts of Lyapunov functions and Bellman's principle of optimality, our agent learns universal state-goal distances and best actions that solve this problem.