Date and LocationNov 06, 2018 - 3:30pm to 4:30pm
AI for Social Good: Decision aids for Countering Terrorism, Extinction, Homelessness
Milind Tambe, University of Southern California
5 November 2018
With the maturing of AI and multiagent systems research, we have a tremendous opportunity to direct these advances towards addressing complex societal problems. I will focus on the problems of countering terrorism (for public safety and security), extinction (wildlife conservation), and homelessness (public health in low resource communities). One key multiagent systems challenge that cuts across these problem areas is how to effectively deploy limited intervention resources. In addressing this challenge, my group has provided novel contributions in multiagent systems research, particularly in terms of computational game theory and agent modelling. For public safety and security, I will introduce our Stackelberg security games model for effectively allocating limited security resources. These security games models are used by agencies such as the US Coast Guard and the US Federal Air Marshals Service to assist in the protection of ports, flights and other critical infrastructure. Second, I will discuss the new green security games to allocate limited resources in protecting endangered wildlife. By advancing adversary modelling in these games, we have helped removal of snares and arrests of poachers in national parks in Uganda. Third, for public health, I will outline the challenges of using limited resources for spreading health information in low resource communities, and algorithms based on games against nature. These algorithms show significant improvements over traditional methods in harnessing social networks to spread HIV-related information among homeless youth. I will also point to directions for future work, illustrating the significant potential of AI for social good.
Finding Efficient Spreaders for Information Diffusion in Social Networks
Bolek Szymanski, Rensselaer Polytechnic Institute
7 November 2018
Recent global events and their poor predictability are often attributed to the complexity of the world event dynamics. To improve the predictability, we use a simple but classic threshold model of contagion spreading in complex social systems in which information propagates with certain probability from nodes just activated to their non-activated neighbors. Diffusion cascades can be triggered by activation of even a small set of nodes. We consider the heterogeneity of individuals' susceptibility to new ideas. We investigate numerically and analytically the transition in the behavior of threshold-limited cascades in the presence of multiple initiators as the distribution of thresholds is varied between the two extreme cases of identical thresholds and a uniform distribution. We show that individuals' heterogeneity of susceptibility governs the dynamics, resulting in different sizes of initiators needed for consensus. Furthermore, given the impact of heterogeneity on the cascade dynamics, we introduce two new selection strategies for Influence Maximization. One of them focuses on finding the balance between targeting nodes which have high resistance to adoptions versus nodes positioned in central spots in networks. The second strategy focuses on the combination of nodes for reaching consensus, by targeting nodes which increase the group's influence. Our strategies outperform other existing strategies regardless of the susceptibility diversity and network degree assortativity. Initial activation of seeds is commonly performed in a single stage. We discuss a novel approach based on sequential seeding. We present experimental results comparing single stage and sequential approaches on directed and undirected graphs to the well-known greedy approach to provide the objective measure of the sequential seeding benefits. Surprisingly, applying sequential seeding to a simple degree-based selection leads to higher coverage than achieved by the expensive greedy approach currently considered the best heuristic.