Vorobeychik Yevgeniy, Joveski Zlatko, Yu Sixie
Electrical Engineering and Computer Science, Vanderbilt University, Nashville, TN, United States of America.
PLoS One. 2017 Feb 8;12(2):e0170780. doi: 10.1371/journal.pone.0170780. eCollection 2017.
Theoretical and experimental investigations have consistently demonstrated that collective performance in a variety of tasks can be significantly improved by allowing communication. We present the results of the first experiment systematically investigating the value of communication in networked consensus. The goal of all tasks in our experiments is for subjects to reach global consensus, even though nodes can only observe choices of their immediate neighbors. Unlike previous networked consensus tasks, our experiments allow subjects to communicate either with their immediate neighbors (locally) or with the entire network (globally). Moreover, we consider treatments in which essentially arbitrary messages can be sent, as well as those in which only one type of message is allowed, informing others about a node's local state. We find that local communication adds minimal value: fraction of games solved is essentially identical to treatments with no communication. Ability to communicate globally, in contrast, offers a significant performance improvement. In addition, we find that constraining people to only exchange messages about local state is significantly better than unconstrained communication. We observe that individual behavior is qualitatively consistent across settings: people clearly react to messages they receive in all communication settings. However, we find that messages received in local communication treatments are relatively uninformative, whereas global communication offers substantial information advantage. Exploring mixed communication settings, in which only a subset of agents are global communicators, we find that a significant number of global communicators is needed for performance to approach success when everyone communicates globally. However, global communicators have a significant advantage: a small tightly connected minority of globally communicating nodes can successfully steer outcomes towards their preferences, although this can be significantly mitigated when all other nodes have the ability to communicate locally with their neighbors.
理论和实验研究一致表明,通过允许交流,各种任务中的集体表现可以得到显著提升。我们展示了首个系统研究网络共识中交流价值的实验结果。我们实验中所有任务的目标是让受试者达成全局共识,尽管节点只能观察其直接邻居的选择。与之前的网络共识任务不同,我们的实验允许受试者与他们的直接邻居(局部地)或与整个网络(全局地)进行交流。此外,我们考虑了可以发送基本任意消息的处理方式,以及只允许发送一种类型消息(即告知其他节点自身局部状态)的处理方式。我们发现局部交流增加的价值微乎其微:解决的博弈比例与无交流的处理方式基本相同。相比之下,全局交流能力带来了显著的性能提升。此外,我们发现限制人们只交流关于局部状态的消息比无限制交流要好得多。我们观察到个体行为在不同设置下在定性上是一致的:人们在所有交流设置中都会对收到的消息做出明显反应。然而,我们发现局部交流处理方式中收到的消息相对信息量不足,而全局交流具有显著的信息优势。在探索混合交流设置(即只有一部分代理是全局交流者)时,我们发现当每个人都进行全局交流时,需要大量的全局交流者才能使性能接近成功。然而,全局交流者具有显著优势:一小部分紧密相连的全局交流节点能够成功地将结果导向他们的偏好,尽管当所有其他节点都有能力与邻居进行局部交流时,这种情况会得到显著缓解。