부산대학교 도서관

Swarm guidance is one of the modern applications of autonomous systems. In it, the controller acts as the transmitter (Tx), and the actuator acts as the receiver (Rx), exchanging control commands through a wireless communication channel to complete the task. The success of the shepherding task, which is a swarm control application, depends on the success of these transmissions through the wireless communication channel that may vary through time due to the Tx's and Rx's mobility. Therefore, to save time and energy, the Tx and Rx must adapt to the communication channel changes while ensuring the task's success in the shortest possible time. In this paper, we model the problem of optimising the mobility of the Tx (controller) to overcome the noise in the commands sent to the Rx (actuator) that acts as a robotic sheepdog in a shepherding scenario with a Markov decision process (MDP). Then, we propose an incremental search method to find near-optimal actions in real-time. Furthermore, two different reward functions are used to help the CU improve its mobility through online learning using Q-learning. The results proved that using Q-learning combined with the incremental search method improved the task's success rate ($SR$) sometimes to exceed the $SR$ in the ideal communication link scenario, which is double its $SR$ in the presence of noise.