Machine learning techniques for autonomous agents in military simulations — Multum in parvo

摘要

In military simulations, software agents are used to represent individuals, weapon platforms or aggregates thereof. Modeling the behavioral capabilities and limitations of such agents may be time-consuming, requiring extensive interaction with subject matter experts and complicated scripts, but nevertheless resulting in rigid, predictable performance. Autonomous agents that learn desired behaviors themselves using Machine Learning (ML) techniques can overcome these shortcomings. However, such techniques are not yet widely used and perhaps underappreciated. In this context, the latin expression 'multum in parvo' ('much in little') denotes that ML agents are able to yield a large variety of behavior, despite their compactness in terms of code and usage of physical memory. This paper attempts to provide some background on applicable Machine Learning solutions and their potential military application. The paper is based on the work of the NATO Research Task Group IST-121 Machine Learning Techniques for Autonomous Computer Generated Entities.