Learned Tactics for Asset Allocation.

摘要

Tactics can be developed in a number of different ways. Rules can be created based on a theory of operations as has been done in the development of tactical decision aids for a considerable time. But these tools can behave poorly in unanticipated scenarios and can require significant design effort. In this paper, an existing machine learning approach for training geographically based agents is used to allocate surveillance assets. While there are many possible approaches to the problem of asset allocation, there are currently few ways of directly comparing these methods for the military environment. It would be advantageous to have a common set of benchmark scenarios that could evaluate different strategies with respect to each other. This paper presents such a problem, simulated data, and a solution. In this domain, two planes (P-3 AIPs) must patrol the waters of Central America with the goal of spotting boats that are smuggling narcotics over a 72-hour period. The area that must be searched is vast and separated by land, so the planes must cooperate to effectively cover it. Of benefit to the planes, however, is that there is some degree of information about the smugglers, although much of it is uncertain. HyperNEAT is a natural fit for this problem because of the large input space and cooperative nature of the problem. The results show that a machine learning approach is able to consistently locate opposing vessels, even in the presence of noise. This approach provides a performance baseline and guide for developing future benchmark problems.