NAVY TECHNOLOGY ACCELERATION (2) - Deep Reinforcement Learning for Decentralized Unmanned Resource Allocation Artificial Intelligence (DURA-AI)

Toyon Research proposes to apply deep reinforcement learning to the problem of managing decentralized teams of unmanned systems to meet complex Navy mission objectives. This research will seek to understand the extent to which an artificial intelligence (AI) can learn to make resource allocation decisions from the experience gained by interacting with a simulated battlespace. First, a simulation-based learning framework will be developed by combining Toyon’s SLAMEM simulator with the open-source projects Gym and TensorFlow. SLAMEM will model a complex mission-of-interest for the Navy and be used to generate training data “episodes” to support the reinforcement learning process. Gym and TensorFlow will deliver a reinforcement learning architecture that supports the parameterization of the AI action policy using a deep neural network. Next, Toyon will design and implement the proof-of-concept Decentralized Unmanned Resource Allocation Artificial Intelligence (DURA-AI) within the new framework. The AI will makes resource allocation decisions in a decentralized architecture to support the mission. The quality of decisions will be quantified by defining a reward function from which the AI will learn good decisions from bad ones. The proof-of-concept AI will be demonstrated and its performance will be measured at the conclusion of the Phase I program.