Connecting Run-Time Metrics to Outcome Performance of Team Attack and Defense

Abstract

Large underactuated ships have been shown to be vulnerable to attacks from smaller, faster boats due to their lack of maneuverability. This scenario can be addressed by a team of small ``satellite'' defensive boats designed to intercept and inspect approaching boats. Movements of defensive agents could be exploited by feinting attackers, drawing aggressive defenders out of an advantageous position as they move to intercept. Moving to defend one region may create vulnerabilities in others. Intelligent strategies are required to perform well in this adversarial environment, and developing these defensive schemes requires effective evaluation of their performance. Unfortunately, defensive strategies are typically evaluated solely by final outcome statistics, which do not carry any information about why a particular outcome was reached. This work seeks to connect outcome performance with run-time metrics that describe the state of a team attack and defense scenario. The teams in this scenario represent the small boats attacking a large ship and the satellite defensive boats used to intercept them. To this end a novel family of run-time metrics, called Time-To-Arrive (TTA), is introduced. TTA can be used by both attack and defense to delegate tasks amongst agents in a team and execute strategies. TTA can be calculated offline to allow for a low run-time complexity. TTA data can be collected for fixed points around the defensive area and manipulated to describe the state of the defense as a whole. Several thousand simulation runs with various combinations of attack and defense strategies, in addition to other varying parameters, were completed and compared. Attackers that exploit TTA information were shown to outperform random attackers. Dynamic defense provides robustness against such exploitation for defenders that can match their opponents' speed, but this turns into a disadvantage for slower defenders. Lower values of TTA at the defensive perimeter showed a strong correlation with higher defensive winning percentage. Lower values of volatility, a scalar measure of how rapidly TTA values changed, also showed a strong correlation with higher defensive winning percentage. These two run-time metrics show a connection to outcome performance, and as such, could serve as a basis for future work as a run-time optimization objective function.