Interactive Augmentation of Computer Generated Force Behavior Based on Cooperative and Reinforcement Learning

State-of-the-art Semi-Automated Forces (SAFs) are predictable, non-adaptive, behavoirally distinguishable from manned simulators, and once "figured out", the only remedy for these performance shortcomimgs is a costly rewrite of the behavioral model software. We propose a hybrid SAF control architecture that integrates neural networks with rule-based systems to enable human like learning, allowing a SAF system to continuously learn from manned simulators, human instructors, and its own mistakes. This technology will make SAF behaviors more realistic, and will save time and money by permiting Battle Trainers to modify SAF behavior without the intervention of programmers. The development team includes experts in robotic skill acquisition and neural network-based control, as well as some of the original developers of SIMNET SAF. Phase I will determine how to integrate concepts of hybrid coorperative and reinforcement learning into an existing SAF system, demonstrate behavior augumentation and performance gains attributable to SAF learning for two battle scenarios, and propose a method for integrating such learning into state-of-the-art ModSAF and Distributed Interactive Simulation.