Leveraging AI for Disaster Preparedness and Response Bistra Dilkina Associate Director, USC Center for AI in Society [CAIS] WiSE Gabilan Assistant Professor, Dept of Computer Science University of Southern California Leveraging AI for Disaster Preparedness and Response Resilient (floods, earthquakes) infrastructure (transportation, water) planning Strategic (positioning, actions) resource (personnel, equipment) allocation LA City/LADWP and Coast Guard logos Preparedness and Response
Critical Infrastructure Systems are Vulnerable to Natural Disasters Given a road network and flood risks, how to upgrade roads to maximize resilience to floods or other disasters? Given water pipe network and earthquake risks, select parts of the network to replace with seismic resilient pipes Critical customers (hospitals, fire/police stations, emergency evacuation centers, power, sanitation, etc) must be directly connected to the resilient network. All households must be within 1mi of the resilient network (reachable by fire hose). Flooded road segments Possible Solutions Number of feasible trips -> related to “normal” functioning of the network, can people get where they need to go Average trip distance -> also related to “normal” functioning, can people get where they need to without having to take long detours Access to hospital -> related to possibly different mobility needs immediately after a disaster Number of alternative routes -> Important for planning evacuations Challenges: limited budget, many subnetworks possible, several resilience metrics Challenges: limited budget, many sub-networks possible, complex constraints
Critical Infrastructure Systems are Vulnerable to Natural Disasters Research Questions: What are the areas of the infrastructure system that are most at risk? How are the consequences of infrastructure disruptions (road, pipes) distributed, who suffers most? How can we decide (near) optimally which upgrades to make maximize resilience, given limited resources? How do we study trade offs between different resilience metrics? Possible Solutions Number of feasible trips -> related to “normal” functioning of the network, can people get where they need to go Average trip distance -> also related to “normal” functioning, can people get where they need to without having to take long detours Access to hospital -> related to possibly different mobility needs immediately after a disaster Number of alternative routes -> Important for planning evacuations
Critical Infrastructure Systems are Vulnerable to Natural Disasters Approach: Large-scale mapping of infrastructure, threats, and human needs Network analysis combined with Predictive Models for flooding/earthquakes Define multiple resilience metrics (through stakeholder meetings) and develop ways to quantitatively assess them under different scenarios Build Pareto frontier of alternative solutions to study systematically trade offs between different resilience metrics Possible Solutions Number of feasible trips -> related to “normal” functioning of the network, can people get where they need to go Average trip distance -> also related to “normal” functioning, can people get where they need to without having to take long detours Access to hospital -> related to possibly different mobility needs immediately after a disaster Number of alternative routes -> Important for planning evacuations
NP-hard optimization problems combinatorial explosion Critical Infrastructure Systems are Vulnerable to Natural Disasters NP-hard optimization problems that suffer from combinatorial explosion Using traditional exact optimization methods (mixed integer program, MIP) Challenge: How can we decide (near) optimally which upgrades to make maximize resilience, given limited resources? ~40 minutes Approach: Develop a portfolio of algorithms including: - optimal baselines - fast heuristic algorithms that work well in practice Leverage new optimization approached augmented with machine learning to speed up solutions over distribution of similar infrastructure problems - Many what-if scenarios to optimize Possible Solutions Number of feasible trips -> related to “normal” functioning of the network, can people get where they need to go Average trip distance -> also related to “normal” functioning, can people get where they need to without having to take long detours Access to hospital -> related to possibly different mobility needs immediately after a disaster Number of alternative routes -> Important for planning evacuations Senegal road network with ~500 road segments would take ~3x1045 years to solve with this approach
Disaster Response Requires Complex and Strategic Coordination Challenge: How agencies can allocate and dynamically reallocate resources/asset teams for effective and timely response, potentially recognizing (additional) resource needs Similar systems already deployed for: anti-poaching ferry patrols TSA screening Approach: To achieve robustness, model arrival of new requests as adversarial Build on hybrid "BDI-POMDP" frameworks to reason about team formation where we assume the presence of some standard operating procedures (SOP), that could be characterized as potentially following a “belief-desire-intention” (BDI) style paradigm. Develop approaches to `mine’ additional resource needs from social media (UIUC) Possible Solutions Number of feasible trips -> related to “normal” functioning of the network, can people get where they need to go Average trip distance -> also related to “normal” functioning, can people get where they need to without having to take long detours Access to hospital -> related to possibly different mobility needs immediately after a disaster Number of alternative routes -> Important for planning evacuations
Overall Benefits to the Department of Homeland Security Natural disasters cost the US 100 billion dollars/year, and the number of people affected continues to grow. Such disasters are likely to increase in frequency, and distribution. Tools and methods for more effective and efficient use of finite assets (funds, personnel teams, equipment) Decision support tools for recommendations and policies for long-term infrastructure investments Ability to systematically study trade offs between performance metrics Broadly applicable computational advancements: Machine Learning-guided planning algorithms that automatically specialize to the optimization problem at hand
Thank You Bistra Dilkina Associate Director, USC Center for AI in Society [CAIS] WiSE Gabilan Assistant Professor, Dept of Computer Science University of Southern California Thank You LA City/LADWP and Coast Guard logos Preparedness and Response