Project

Smart Emergency Response - Integrated Safety Incident Forecasting and Analysis

The objective of this research is to understand and improve the resource coordination and dispatch mechanisms used by first responders. The problem of dispatching emergency responders to service accidents, fire, distress calls and crimes plagues urban areas across the globe. In prior art, as well as practice, incident forecasting and response are typically siloed by category and department, reducing effectiveness of prediction and precluding efficient coordination of resources. Further, most of these approaches are offline and fail to capture the dynamically changing environments under which critical emergency response occurs, and therefore, fail to be implemented in practice. Consider the classical problem of emergency response. The goal of responders is to minimize the variance in the operational delay between the time incidents are reported and when responders arrive on the scene.

Solving this problem requires not just sending the nearest emergency responder, but sometimes being proactive placing emergency vehicles in regions with higher incident likelihood. Sending the nearest available responder by euclidean distance ignores road networks and their congestion, as well as where the resources are stationed. Greedily assigning resources to incidents can lead to resources being pulled away from their stations, increasing response times if an incident occurs in the future in the area where responder should be positioned. Now, consider solving this problem when there is a high uncertainty in the veracity of the request due to either communication failures or due to the nature of the communication medium – in extreme disruptions the most common communication mechanism used is social media, however, the social media requests have a lot of uncertainty in terms of duplication, spatial location etc. Ultimately, the methods developed in this work can be applied to other domains where multi-resource spatio-temporal scheduling is a challenge. We collaborate with Prof. Yevgeniy Vorobeychik, WUSTL, Prof. Hemant Purohit, GMU and Prof. Saideep Nannapaneni, Wichita State.

Any holistic approach towards creating a pipeline for effective emergency response must also look at other challenges that it subsumes - predicting when and where incidents happen and understanding the changing environmental dynamics and possible communication disruptions. This research project provides a unique opportunity to study these problem by integrating both the data and emergency resources from distinct urban agencies in the City of Nashville along with other widely available data such as pedestrian traffic, road characteristics, traffic congestion, and weather. This will allow development of models for anticipating heterogeneous incidents, such as distinct categories of crime, as well as vehicular accidents. With these models we can develop decision support tools to optimize both resource allocation and response times. These tools will help the emergency responders determine which units to dispatch (police, fire, or both) in order to minimize expected response time, and what equipment is most appropriate, taking into account the time, location, and nature of incidents, as well as those predicted to occur in the future. Though funding from the National Science Foundation, our team has been working on five principled approaches to solve this problem in collaboration with the city of Nashville.

  1. Incident Prediction Using Online Survival Analysis and Long Short Term Memory Networks - We have developed a novel online approach to incident prediction that predicts incidents in time and space. Previous work in this domain has treated this as a batch learning problem in which incident prediction models are learned once, and are subsequently used to aid response decisions. This fails to capture the changing dynamics of urban systems in which emergency responders operate, and we bridge this gap by creating an online incident prediction algorithm. Our framework includes an online survival model for incident prediction and a recurrent neural network model for learning environmental features affecting dispatch.
  2. Uncertainty quantification by Uncertain Concept Graphs (UCG)- We have been developing the theory of uncertain concept graphs that combine graphical fault propagation models with dynamic Bayesian networks. The UCG is capable of representing dynamic knowledge of a disaster event from heterogeneous data sources, particularly for the regions of interest, and resources/services required. The information sources, incident regions, and resources (e.g., ambulances) are represented as nodes in UCG, while the edges represent the weighted relationships between these nodes. We have developed a theoretical solution for probabilistic edge inference between nodes in UCG. The output of such structured summarization over time can be valuable for modeling event dynamics for the decision support in the real world beyond emergency management, across different smart city operations such as transportation.
  3. Dispatch Suggestion Framework - We formulate the problem of dispatching responders to incidents as a Semi Markov Decision Process (SMDP). However, solving this class of problems online is extremely slow and fails to work in dynamic environments since any change in the problem definition (the number of responders, or the position of a depot) renders the learned policy stale. In order to tackle this problem, we use an important observation - one need not find an optimal action for each state as part of the solution approach since at any point in time, only one decision-making state might arise that requires an optimal action. This difference is crucial, as it lets us bypass the need to learn an optimal policy for the entire MDP. Instead, we describe a principled approach that evaluates different actions at a given state, and selects the one that is sufficiently close to the optimal action. We do this using sparse sampling, which creates a sub-MDP around the neighborhood of the given state and then searches that neighborhood for an action. In order to actualize this, we use Monte-Carlo Tree Search (MCTS).
  4. Policy Framework For Evaluation of Decisions - We have been developing a framework for testing the dispatch algorithms and then enable reasoning over anomalous decisions. Specifically, to incorporate the future environment in dispatch decisions, a discrete event simulator is used. It consists of a grid-based model of the environment, including where EMS/Fire stations are located. For each grid cell, there is a learned incident prediction model (using survival analysis) that is used to sample future likely incidents. Responders and their states are represented by agents that move around the grid from stations to incidents, and a traffic model and router are used to simulate travel times between grids. When an incident occurs and a dispatching decision must be made, the following steps occur. First, future incidents are sampled from the spatio-temporal prediction models. Then, each agent builds a Monte Carlo Search Tree to estimate the utility of each of its potential actions. Future incidents are used as future states, and the actions of other agents are approximated.
  5. Resilient Distributed Middleware- The online decision framework we have built is only as good as the communication framework and the computation framework on which it can run. In extreme circumstances the whole city can be cut off from the cloud providers and might not be in a position to run the online framework. To support partitioned execution capability we have been developing a decentralized middleware that is capable of recovering from communication network partitions and distributed computations tasks across available edge computation resources. The key to resilience in this framework is the use of a distributed ledger to maintain data consistency and use marked based mechanisms to offload tasks to computation resources. A smart contract in the system is responsible for implementing the task placement correctly.