Online Cognitive MFR Task Scheduling Using MCTS and a Trained Model

Task scheduling in a cognitive multifunction radar (MFR) must adapt to continuously changing operating environments. Recent radar task scheduling approaches assume knowledge of the environment, which is unknown in real scenarios. Instead of relying on assumed knowledge, this work investigates the online model-based deep reinforcement learning (DRL), specifically MuZero, which learns model of the environment to optimally plan for scheduling the tasks. In this approach, the scheduler learns an abstract Markov decision process (MDP) model of the environment, allowing near-optimal planning in the abstract MDP to translate effectively into the real environment. However, the original MuZero approach considers a model with three deep neural networks, which require very long training times in the MFR task scheduling problem, specifically for larger number of tasks with different priorities. We modify the original MuZero algorithm to accommodate large action spaces by incorporating prior knowledge of the task scheduling problem, thus requiring only a policy network to train. Our numerical results show that the modified MuZero approach is effective and computationally efficient in complex radar scenarios.