On improving the properties of random walk on graph using Q-learning

Abstract

Studies have recently been conducted to model mobile agents on unknown graphs, such as random walks (RWs) on graphs, and to understand their mathematical properties. In this study, we investigate the extent to which the properties of RWs can be improved when mobile agents have access to very limited information. We propose Q-weighted random walk (QW-RW), in which an agent decides a next node by using Q-values learned by Q-learning, and examine its effectiveness. We find that in small scale-free graphs, QW-RW is 1.25 times faster than self-avoiding RW to cover 80% of the entire graph.