Quick Look
Using Pebble® Decision Models for
Speech Recognition/Parsing
and for Replacement of Human Operators in Training
and Simulation Platforms, 07-9336
Principal Investigators
Peter Hottenstein
Tom Glass
Inclusive Dates: 07/11/02 - 11/11/02
Background - Most simulators used by the military for training require many people working behind the scenes to generate a realistic simulated battle space for a single student. Many attempts have been made to develop sophisticated artificial intelligence agents to control these simulations, but the complexity of true artificial intelligence has made them impractical for use in day-to-day training. The lack of predictable results is not conducive to a training environment.
Approach - The research team's goal was to develop a demonstration of how a simpler decision model can be incorporated into the Institute-developed AWACS Modeling and Simulation (AMS) training system that is used to train U.S. Air Force Air Battle Managers. This demonstration proved the utility and versatility of using a predictable and maintainable decision modeling tool such as the Parametric Engine for Building Boolean Logic Entities™ (PEBBLE®). PEBBLE, a software application developed under a previous SwRI internal research project, is an easy to maintain decision-modeling application. The research team developed a detailed training scenario and the associated logic for several decision models. An interface was designed to allow real-time communications between the Java based PEBBLE and the C++ based AMS system, enabling operators to update the decision models without requiring the training system software to be recompiled. A speech recognition input capability was added to the PEBBLE interface so that decisions within the model are based on dynamic operator input as well as predefined environmental factors.
Accomplishments - The research team successfully developed a demonstration of the capabilities of PEBBLE in the AMS simulation for an airborne interdiction training scenario. As part of this effort, the team built and integrated six decision models representing human decisions of different pilots with varying levels of proficiency. The logic used for these decision models has been tested and exhibits a level of behavioral competence with enough fidelity to ensure realistic training.
