Intelligent Counter Guidance Regulated by Deep Reinforced Learning

In the game of pursuit and evasion in outer space, it is of utmost importance for the evader to make appropriate maneuver to survive this challenge. In previous publications, the common approach of guidance behavior recognition and counteracting evasion approach is based on model matching or multi-model matching. In that genre, due to the limitation of approximation from model matching and imperfection of state information, it is very difficult for the evader to take optimal maneuver when the pursuer uses unknown tracking algorithms. To solve this problem, in this article, deep neural network and reinforced learning are applied in counteractive maneuver. In the proposed approach, two networks are applied. Guidance recognition network is a pre-trained deep network able to estimate guidance behavior of pursuer vehicle, and the maneuver control network is trained by deep reinforced learning to perform counteract evading maneuver control. Exploring the policy space with Monte Carlo method, the maneuver controller deep neural network evolved to make optimal control acceleration and succeed in the evasion. Digital simulations confirmed that in the game of pursuit and evasion, deep reinforced learning with Monte Carlo method has the ability to guide the evader during the interception with improved chance of survival.