Critical to Army readiness, simulation-based training offers a cost- and time-effective way to keep personnel well-versed in their roles, responsibilities, tactics, and operations. Simulation supported exercises currently require long planning timelines and significant facility and personnel resources. Though semi-automated military simulations provide basic behavioral artificial intelligence to assist in fulfilling participant roles, they still require human simulation operators to control friendly and opposing forces. Often these exercise support simulation operators come directly from the intended training audience assigning Soldiers role-playing duties versus training with their organization. Simulations are intended to make the training process simple and repeatable. However, the current overhead of simulation supported training events overwhelm units and prevent them from getting quality training repetitions needed to gain proficiency. One method of reducing this overhead is through machine learning trained, adaptive opposition forces (OPFOR).
Recent machine learning advancements, such as Deep Mind’s AlphaZero and AlphaStar, produced computerized agents capable of defeating professional human players in complex strategy games in real-time. Complex strategy games create an environment consisting of multiple agents capable of making their own decisions while sharing a common goal. Furthermore, these multi-agent platforms typically present imperfect information for the agents due to fog of war, contain large state–spaces, and require three-dimensional terrain navigation. These feats help support the idea that machine learning may be the key to developing adaptive OPFOR in military constructive simulations.
This paper surveys the existing literature in the use of machine learning for automated OPFOR decision making, plan classification, and agent coordination. Further, this paper reviews current advancements in temporal-adaptation to analyze how quickly machine learning agents can adapt to a change in their environment. This analysis serves as a starting point for any future research in the current capabilities and limitations of developing adaptive OPFOR in support of military constructive simulations.
Keywords
ADAPTIVE,COMPUTER GENERATED FORCES,MACHINE LEARNING,REAL-TIME,SIMULATIONS
Additional Keywords