LVC Simulation for Improved Training Efficiency

Important results from the project

The project has studied how Live, Virtual, and Constructive (LVC) simulations can be combined to provide efficient and effective pilot training. The focus has been on development of methods based on machine learning, for automation of tasks that otherwise would be performed by support personnel, such as role-players, instructors, and scenario operators. The project has established a knowledgebase at low TRL, as well as collaborations and contacts with national and international actors within related areas, which can be used to support future research and development.

Expected long term effects

The project has focused on the machine learning technique reinforcement learning, and scenarios where groups of agents must learn how to solve complex problems in environments where multiple, possibly conflicting objectives must be considered, e.g., tactical mission goals, resource consumption and safety. User studies have identified important aspects of synthetic training environments. The research has resulted in several publications, and collaborations between industry, academia, and the air force have ensured that results have been spread for maximal effect.

Approach and implementation

The project was divided into two phases: Phase one conducted a literature survey and initial analysis and evaluation of existing state-of-the-art within reinforcement learning, and the second phase focused on further development and adaptation of algorithms for the domain of air combat training. Throughout the project, interaction with pilots was used for evaluation of results. Identified challenges within the air combat domain have resulted in that the project, compared to the original plan, has focused more on algorithm development than on evaluations with pilots in the loop.