Creating and Evaluating Goal Ordering Structures for Testing Harbour Patrol and Interception Policies

Abstract

In this article, we discuss a method for testing policies that guide groups of agents in simulations for interactions with other agents and the environment that reveal weaknesses of these policies. Our method is based on learning interaction sequences using particle swarm systems and has as one crucial component so-called goal ordering structures that are used to guide the learning towards weakness-revealing interactions. Our discussion centers around the different ways a new measuring idea can be integrated into such an ordering structure using the example of testing patrol and interception policies for harbours. Our experimental evaluation reveals that the position of placement of a new measure in an existing ordering structure can greatly influence the testing results, positively and negatively, but mostly mirrors the intuition associated with the placement.