Safe navigation in adversarial environments

This work deals with the problem of navigation while avoiding detection by a mobile adversary, featuring adversarial modeling. In this problem, an evading agent is placed on a graph, where one or more nodes are defined as safehouses. The agent’s goal is to find a path from its current location to a safehouse, while minimizing the probability of meeting a mobile adversarial agent at a node along its path (i.e., being captured). We examine several models of this problem, where each one has different assumptions on what the agents know about their opponent, all using a framework for computing node utility, introduced herein. Using risk attitudes for computing the utility values, their impact on the constructed strategies is analyzed both theoretically and empirically. Furthermore, we allow the agents to use information gained along their movement, in order to efficiently update their motion strategies on-the-fly. Theoretical and empirical analysis shows the importance of using this information and these on-the-fly strategy updates.