Decision support in regulating decentralized multi-agent systems

Abstract

A common challenge in modern real-world systems, such as transportation system, building management system and financial market, is dynamic resource allocation in decentralized multi-agent systems. In such scenarios, self-interested autonomous agents compete for limited resources. A regulator, who usually cannot observe many of the agents’ preferences, designs and implements interventions in real time in attempt to bring about societal goals. Both the welfare of individuals and society as a whole must be satisfied to sustain a successful system. In this work, we developed a forecasting tool designed to help regulators set effective interventions under various settings. We then conducted a user study (using Jiao Tong and 48 human subjects) to empirically evaluate how this forecasting tool would impact the performance of the regulatory entity in optimizing the aggregate societal welfare in real-time interventions. Results from the user study indicated that this forecasting tool did not help regulator’s to improve the system’s objective performance. Rather, the forecasting tool distracted the regulators from understanding and modeling the problem correctly. The results of the user study clearly demonstrated that the forecasting tool should be improved. Although it helped draw users’ attentions to the future hazards (congestions), it failed to indicate what measures the regulators could take to avoid or mitigate the hazards. Therefore, these results advocate that forecasting tools for such systems must both alert regulators of potential hazards and provide directions for how the problem should be mitigated.

Date of Award	Dec 2014
Original language	American English
Supervisor	Jacob Crandall (Supervisor)