In order for simulation based training to help prepare soldiers for modern asymmetric tactics, opponent models of behavior must become more dynamic and challenge trainees with adaptive threats consistent with those encountered increasingly in the real world. In this presentation we describe an adaptive behavior modeling framework designed to represent adversaries within a multi-player virtual environment. Two distinct areas of investigation are covered. The first area is a survey of the space of asymmetric tactics and adaptations from real-world military operations, to generate a set of reference scenarios. The second research area, and the focus of this presentation, is the design and development of machine learning techniques for creating adaptive adversaries. The approach makes use of an authoring tool for specifying adaptive behavior models as partial plans that can adapt over time in conjunction with training events. This approach focuses on supporting both a natural method of encoding existing domain knowledge and the rapid adaptation of encoded behaviors. The overall objective for this approach is that the adversary behavior models should constantly challenge, and occasionally surprise, the human trainees, to help them learn to be more proactive in recognizing asymmetric threats.