Study on a Resilient Helping Role Based on Top-down and Bottom-up Processing in Coordinated Behavior of a Triad

抄録

A group distributes roles to achieve a common goal, enabling higher task performance than when doing alone. Such coordination has been investigated in various research fields. These findings suggest that two types of information processing work for efficient and adaptable behaviors: (1) top-down processing established by structured internal knowledge and representation of a group goal, task constraints, plans, and roles, and (2) bottom-up processing based on sensory inputs, such as the generation of flexible movement itself through perception. However, coordination mechanisms have not been fully discussed in terms of the two types of processing. Meanwhile, a previous cognitive science study identified a crucial role for coordination. In a coordinated drawing task, a participant triad shares heterogeneous roles and changes each tension using a reel to move a pen connected to three threads to draw an equilateral triangle. The results indicated that a resilient helping role, which moderately intervenes with other roles to adjust the whole balance according to situations, was related to high team performance. Although this role is not only required for the experimental task, it has not been explained in related work. Considering the aforementioned discussions, the adjustment process particularly involves the two types of processing; however, there is room for further investigation. This study introduced computer simulation to the coordinated drawing task and examined the resilient helping role, using deep reinforcement learning and rule-based modeling. The results showed that an agent with an interactive relationship model in which top-down processing drives bottom-up processing was able to adjust and correct the pen trajectory at the proper timing. Additionally, the deep reinforcement learning and rule-based condition in the adjusting role achieved higher team performance (smaller pen deviation) than the rule-based alone and random conditions. This study supplements the experimental findings and contributes to a constructive understanding of coordination.