Reinforcement Learning Based Decentralized Weapon-Target Assignment and Guidance

Multiple-missile attack is one of the simplest ways to saturate and overcome a missile defense system. To increase intercept efficiency against such groups of attackers, it is necessary to allocate the interceptors according to their kinematic limitations. Moreover, such an allocation scheme has to be scalable in order to cope with large scenarios and allow for dynamic reallocation. In this paper we first propose a new formulation of such a Weapon-Target Assignment (WTA) problem and offer a decentralized approach to solve it using Reinforcement Learning (RL) as well as a greedy search algorithm. The engagement is considered from the viewpoint of each pursuer vs. all the targets. Simultaneously, other interceptors engage the target group, and their allocation and success probabilities are available to other team members. To improve the mid-course trajectory shaping, static virtual targets are placed between the pursuers and the incoming adversaries. Each interceptor selects its target dynamically according to a policy that was learned from a large number of scenarios in the computation-efficient simulation environment. The RL input state contains the interceptor reachability coverage of the targets and the probabilities of success of other missiles. The RL reward aggregates the team performance to encourage cooperation on the allocation level. The relevant reachability constraints are obtained analytically by employing kinematic approximations of the interceptor motion. The use of RL ensures real-time scalable and dynamic reallocation for all interceptors. We compare the performance of the proposed RL-based decentralized WTA and guidance scheme against a greedy solution, showing the performance advantage of RL.