Event-Triggered Deep Reinforcement Learning for Dynamic Task Scheduling in Multisatellite Resource Allocation

In this work, we investigate the problem of multisatellite resource allocation for expected long-term performance optimization with a dynamic task network model, where communication tasks generated by task satellites are expected to be transmitted by resource satellites in the application layer, and the set of tasks changes with satellite orbital motions. The features of the tasks include priority, execution duration, visible time, etc. Since the feature information has a high dimension and changes with time, the scheduling problem is formulated as a dynamic combinatorial optimization problem and a receding-horizon task scheduling algorithm based on the event-triggered deep reinforcement learning is proposed. A residual-fully connected network is designed to extract the features of the complex task network model, and a deep double Q-learning iteration with the experience replay memory mechanism is employed to change the allocation strategy by evaluated rewards adaptively. An event-triggered strategy is then proposed to handle urgent tasks online. Numerical simulations show the performance improvement of the proposed algorithm. For the scenario of 50 task satellites and ten resource satellites, the proposed algorithm achieves 4.1%, 5.9%, and 11.4% higher reward scores than the static deep reinforcement learning algorithm, the data-driven parallel scheduling algorithm, and the improved genetic algorithm, respectively. The computation time of the proposed algorithm is only 34.7% and 21.3% of that of the latter two algorithms, and is similar to that of the static deep reinforcement learning algorithm.