(22c) The Control of Self-Interested Agents: Learning from Nature’s Wisdom of Crowds

Multi-agent collective dynamics is a prevalent phenomenon. Examples include self-assembly, coupled oscillators, cell migration, quorum sensing, birds flocking, ant colony, etc., which all share something in common: Individuals interact with the crowds that they belong to and benefit from such interactions. â??The wisdom of crowds,â? a century-old discovery in statistics, offers an intuitive explanation: The average opinion formed by a large population can be remarkably accurate. However, supporting theories assume that individual opinions are independent and unbiased. Interactions would thus undermine the wisdom. Those theories, therefore, fail to explain why interacting agents in nature still choose to rely on such wisdom.

In this talk, we formulate this phenomenon as a standard control problem and identify several properties overlooked by existing theories. This work explains the wisdom of interacting crowds and sheds light in achieving the elusive goal of controlling self-interested agents.

Our model consists of two processes: First, the agents are self-interested and utility-maximizing. They follow an internal optimization mechanism to achieve the goals of survival or proliferation. In the cell migration system, for example, each cell responds to the concentration gradient signals and adjusts its motion for self-preservation. Sensory noise and cognitive limitation might disturb such process. Second, the agents partially follow the signals from their neighbors, the quorum, or simply the mean-field. We formulate those signals as the control input, which is "softer" than what one might observe in process control. We name such control model "soft regulation."

Through mathematical proofs, simulations, and Amazon Mechanical Turk experiments with human subjects, we show that soft regulation can improve both the individual and the social welfares from the open loop. This result thus agrees with what we observe from nature without restricting us to the condition that requires agents to be independent. Furthermore, we identify how uncertainty, individual heterogeneity, social influence, and network structure affect the efficacy of this mechanism. Such characterizations will help us design and improve control mechanisms for sociotechnical systems.

After all, just like flocking birds, human individuals, organizations, and even nations are also crowds of self-interested agents with common goals of survival and proliferation. Those goals may be improving health and fitness, regulating emerging technologies, or fighting climate change. Learning from natureâ??s wisdom of crowds is the first step of achieving the control of self-interested agents.