Statistical Optimization and Analytics for Community Emergency Management
Context: The goal of this project is to improve emergency response systems using proactive resource management that minimizes time and maximizes the effectiveness of the response. With road accidents accounting for 1.25 million deaths globally and 240 million emergency medical services (EMS) calls in the U.S. each year, there is a critical need for a proactive and effective response to these emergencies. Furthermore, a timely response to these incidents is crucial and life-saving for severe incidents. The process of managing emergencies requires full integration of planning and response data and models and their implementation in a dynamic and uncertain environment to support real-time decisions of dispatching emergency response resources. However, the current state-of-the-art research has mainly focused on advances that target individual aspects of emergency response (e.g., prediction, optimization) when different components of an Emergency Response Management (ERM) system are highly interconnected. Additionally, the current practice of ERM workflow in the U.S. is reactive, resulting in a large variance in response times
Background: Our work in ERM has spanned the last six years. This project was started by a collaboration between the Smart and Resilient Computing for Physical Environments Lab (SCOPE) and the Computational Economics Research Lab (CERL) at Vanderbilt University and is currently developed jointly by the Stanford Intelligent Systems Lab (SISL) at Stanford and the SCOPE Lab. We are thankful to the Center of Automotive Research at Stanford (CARS), the National Science Foundation (NSF), and the Tennessee Department of Transportation (TDOT) for sponsoring the project. We have had the fortune of collaborating with the Tennessee Department of Transportation (TDOT), the Nashville Fire Department (MNPD), and Chattanooga City during this project. Currently, this open-source repository is a collection of forecasting, planning, and operationalization tools. We also collaborate actively with Hemant Purohit from George Mason University to model the dynamics of crowdsourced incident data and use it in the resource allocation models. A key component of this work is a set of open-source forecasting, clustering, and visualization tools to aid first responders better understand the dynamics of spatial-temporal incident occurrence.
Innovation: We use continuous-time generative models to forecast spatiotemporal incidents and the decision-theoretic problem of dispatching responders based on semi-Markovian dynamics. We have also developed efficient and scalable approaches to solve the high-dimensional optimization problem of proactive stationing and dispatch under uncertainty by using Multi-agent Monte Carlo Tree Search (MMCTS). This is important because the problem of stationing and dispatch (response strategies) is complex and requires solutions that can cope with extremely large state spaces resulting from semi-Markovian decision processes (SMDP). One possible approach is to directly solve the SMDP model by estimating the underlying transition function that governs the evolution of the process. Unfortunately, this approach is too slow for dynamically rebalancing the distribution of responders (proactive stationing), which for an average-sized metropolitan area, has a cardinality of 10^25. Therefore, we use the Monte-Carlo Tree Search (MCTS) family of algorithms, which evaluate actions by sampling from a large number of possible scenarios.
Note that a standard MCTS-based approach is not suitable for dynamic allocation due to the sheer size of the state-space in consideration coupled with the low latency that ERM systems can afford. Therefore, we sought a decentralized multi-agent MCTS (MMCTS) approach, originally explored by Claes et. al for multi-robot task allocation during warehouse commissioning. In MMCTS, individual agents build separate trees focused on their own actions, rather than having one monolithic, centralized tree, dramatically reducing their search space. To this end, we use a queue-based rebalancing heuristic (described in our AAMAS 2020 paper) to approximate agent behavior. Our innovation lies in adding the concept of action filtering to the standard MCTS approach. This is required because the dispatching domain has several global constraints to adhere to, such as ensuring that an incident is serviced if possible. Figures c and d show the results from the AAMAS 2020 paper indicating that proper choice of parameters reduces response time variance without incurring large rebalancing costs (i.e. the distance agents move during rebalancing). Similar improvements accrue for the dispatching stage as shown in our ICCPS paper. Using these methods we have also developed analysis procedures to help answer questions such as
where should the next fire station be located'',how many new emergency dispatch trucks are required’’ and reducing the average response time as shown in our existing research.
A key aspect of our solution approach is development of demand forecasting models that describe the need for resources in a given area at a given future time. For this we build forecasting models using historical incident data, temporal data (weather and traffic) and static roadway data. The data is processed and fed into an aggregator, which identifies clusters of roadway segments (or any user-specified unit of spatial discretization) that are similar to each other using a visual analytic tool. For each such cluster, a set of online forecasting models are learned and compared automatically using criteria such as test-set likelihood and AIC (Akaike Information Criteria), and are updated as new incidents are reported.Our tool currently supports Poisson regression, negative-binomial regression, parametric survival modeling and zero-inflated Poisson regression.
Media: Our research has been showcased at multiple global smart city summits, won an innovation from the government technology magazine written by Metrolabs, covered in the Financial Times, and won the best paper award at ICLR’s AI for Social Good Workshop. Our broader approach can be understood from checking the Research page, or through the overview paper.
Further Information is available at statresp.ai