Task Scheduling in Cognitive Multifunction Radar Using Model-Based DRL

There has been much recent work on machine learning-based approaches for cognitive task scheduling in multifunction radar (MFR). However, the available MFR scheduling approaches rely on knowledge of the operating environment; in practice, though, the inherent uncertainty with dynamic radar environments poses significant challenges, especially for cognitive MFR. Here, we address the need for online task scheduling in a cognitive MFR without prior knowledge of the environment. We draw inspiration from recent advancements in model-based deep reinforcement learning, specifically the applicability of MuZero, to enable cognitive MFR in unknown and continuously changing environments. In this approach, the scheduler learns an abstract Markov decision process (MDP) model of the environment, allowing near-optimal abstract MDP planning to translate effectively into the actual environment. However, with exponential complexity, the published MuZero approach requires long training times and is useful only for a few tasks. Here, we modify the original MuZero algorithm to accommodate large number of tasks by incorporating prior knowledge of the task scheduling problem. Our numerical results show that the modified-MuZero approach is effective and computationally efficient in complex radar scenarios.