Goal-directed graph construction using reinforcement learning

Graphs can be used to represent and reason about systems a variety of metrics have been devised quantify their global characteristics. However, little is currently known how construct graph or improve an existing one given target objective. In this work, we formulate the construction as decision-making process in which central agent creates topologies by trial error receives rewards proportional value By means conceptual framework, propose algorithm based on reinforcement learning neural networks learn improvement strategies. Our core case study focuses robustness failures attacks, property relevant for infrastructure communication that power modern society. Experiments synthetic real-world graphs show approach outperform methods while being cheaper evaluate. It also allows generalization out-of-sample graphs, well larger out-of-distribution some cases. The applicable optimization other structural properties graphs.