視野領域を制限した追跡問題におけるProfit Sharing法に基づく学習方法の提案

抄録

To study security at facilities, particularly the capture of a suspicious person using only security robots, we investigated the pursuit problem. Multi-agent reinforcement learning has frequently been applied to this problem, in which multiple hunters catch single or multiple preys. In the pursuit problem, hunters are typically considered to have an infinite field of view (FoV) range to obtain the absolute or relative positions of the preys as the states of the hunters. However, this is inappropriate owing to “the curse of dimensionality,” and the FoV range should be restricted suitably. Moreover, strict restrictions on the FoV range prevent hunters from observing their states and continuing learning because they cannot obtain the position of a prey when it is outside their FoV. In previous studies, when a prey is out of the FoV of the hunters, they act randomly. However, in this study, we developed methods to allow hunters to continue learning when a prey is outside their FoV. In addition, numerical simulations were performed to evaluate the effectiveness of each method. The simulation results validated the effectiveness of developed methods from the viewpoints of learning performance and application of the learning results.