Visualizations for Communicating Intelligent Agent Generated Courses of Action

Abstract

Future human-autonomy teams will benefit from intelligent agents that can quickly deliberate across multiple parameters to generate candidate courses of action (COAs). This experiment evaluated the design of an interface to communicate agent-generated COAs to a human operator. Twelve participants completed 14 trials, each consisting of a series of tasks that required participants’ selection of the best COA in terms of quality, speed, fuel, and detectability parameters. Trial score and speed of participants’ selection were measured as a function of COA visualization (1, 4, or 8 COAs) as well as the type of agent. Supplemental trials in which participants could choose which visualization to employ for COA selection were also conducted. The data showed that presenting multiple COAs were better than a single COA. Differences between the 4 and 8 COA visualizations were not quite as definitive: selections were significantly faster with 4 COAs than 8, but participants’ preferences were divided based upon agent comprehensiveness and individual strategy differences. The results also showed that the agent’s reasoning process should be communicated more precisely besides just what parameters are being considered in generating COAs.