SOAR TECHNOLOGY, LLC — Department of Defense SBIR Phase I: N171-084

Description

SoarTech proposes DeepAgent, a system for automatically learning and improving multiple, realistic behaviors in diverse, complex environments using observations of human gameplay. DeepAgent instruments a training simulation to record a compact representation of the simulation state, player actions, and reward at every time step. Deep reinforcement learning is used to learn deep neural networks which are specially designed to approximate value and policy functions in complex environments. At this point DeepAgent can begin to control agents in the simulation by following the learned policy function. This entire process can be repeated using additional observations or by having DeepAgent play against itself, leading to continuous improvement in the behavior.