Planning spatial networks with Monte Carlo tree search

We tackle the problem of goal-directed graph construction: given a starting graph, finding set edges whose addition maximally improves global objective function. This emerges in many transportation and infrastructure networks that are critical importance to society. identify two significant shortcomings present reinforcement learning methods: their exclusive focus on topology detriment spatial characteristics (which known influence growth density links), as well rapid action spaces costs model training. Our formulation deterministic Markov decision process allows us adopt Monte Carlo tree search framework, an artificial intelligence decision-time planning method. propose improvements over standard upper confidence bounds for trees (UCT) algorithm this family problems addresses single-agent nature, trade-off between cost contribution objective, space linear number nodes. approach yields substantial UCT increasing efficiency attack resilience synthetic real-world Internet backbone metro systems, while using wall clock time budget similar other search-based algorithms. also demonstrate our scales significantly larger than previous methods, since it does not require training model.