Reinforcement-Learning-Based Cooperative Dynamic Weapon-Target Assignment in a Multiagent Engagement

This paper considers a multiagent Shoot-Shoot-Look engagement scenario, in which multiple pursuers act against multiple evaders with predefined motion. The pursuers are arranged in two successive waves - the first wave engages the a-priori allocated evaders directly, and the second wave trails behind to assist with the pursuit of the evaders that survive the firstwave. The scenario objective is to maximize the number of intercepted evaders under the given time constraints. To facilitate the intercept performance, the second-wave pursuers are guided to intermediate virtual targets that allow several reallocation options when the actual first-wave engagement outcomes are known. The choice of the virtual targets influences the time and maneuver required from the pursuer during the engagement, which are related to the intercept probabilities. We formulate the problem of second-wave allocation as a stochastic Markov Decision Process. Due to the special problem structure, a reward expression based on the predictions of intercept probabilities is developed. Using this reward function, a Reinforcement-Learning-based strategy is proposed for the virtual target allocation for the second-wave pursuers. An alternative Greedy algorithm is designed based on maximizing individual contribution increments into intercept probabilities. Sequential decentralized decision-making architecture is used to implement both approaches. The simulation demonstrate that the Reinforcement-Learning-based solution wins slightly over Greedy, and the proposed greedy heuristic approximates well the Reinforcement-Learning solution.