=Paper=
{{Paper
|id=Vol-2884/paper_105
|storemode=property
|title=Designing Emergency Response Pipelines : Lessons and Challenges
|pdfUrl=https://ceur-ws.org/Vol-2884/paper_105.pdf
|volume=Vol-2884
|authors=Ayan Mukhopadhyay,Geoffrey Pettet,Mykel Kochenderfer,Abhishek Dubey
|dblpUrl=https://dblp.org/rec/conf/aaaifs/MukhopadhyayPKD20
}}
==Designing Emergency Response Pipelines : Lessons and Challenges==
https://ceur-ws.org/Vol-2884/paper_105.pdf
Designing Emergency Response Pipelines : Lessons and Challenges
Ayan Mukhopadhyay, 1 Geoffrey Pettet, 2 Mykel Kochenderfer, 1 Abhishek Dubey 2
1
Stanford University,
2
Vanderbilt University
1
{ayanmukh, mykel}@stanford.edu, 2 {abhishek.dubey, geoffrey.a.pettet}@vanderbilt.edu
Abstract tained efforts to reduce long-term risks to people and prop-
erty. It also involves creating forecasting models to un-
Emergency response to incidents such as accidents, crimes,
derstand the spatial and temporal characteristics of inci-
and fires is a major problem faced by communities. Emer-
gency response management comprises of several stages and dents. Preparedness involves creating policy and allocating
sub-problems like forecasting, resource allocation, and dis- resources that enable emergency response management. The
patch. The design of principled approaches to tackle each third phase seeks to use automated techniques to detect in-
problem is necessary to create efficient emergency response cidents as they happen in order to expedite response. The
management (ERM) pipelines. Over the last six years, we dispatch phase, the most critical phase in the field, involves
have worked with several first responder organizations to responding to incidents when they occur. Finally, the recov-
design ERM pipelines. In this paper, we highlight some of ery phase seeks to provide support and sustenance to com-
the challenges that we have identified and lessons that we munities and individuals affected by the emergency. While
have learned through our experience in this domain. Such most of the prior work in ERM has studied these problems
challenges are particularly relevant for practitioners and re-
independently, these stages are inter-linked. Frequently, the
searchers, and are important considerations even in the de-
sign of response strategies to mitigate disasters like floods output of one stage serves as the input for another. For exam-
and earthquakes. ple, predictive models learned in the preparedness stage are
used in planning response strategies. Therefore, it is crucial
that ERM pipelines are designed such that intricate interde-
Introduction pendencies are considered. In this paper, we highlight some
Emergency response to incidents such as accidents, medical of the challenges that we have faced while designing ERM
calls, urban crimes, poaching, and fires is one of the most pipelines and lessons that we have learned through the pro-
pressing problems faced by communities across the globe. cess. A large portion of the insights have come from our col-
Emergency response must be fast and efficient in order to leagues at the Nashville Fire Department and the Tennessee
minimize loss of lives (Jaldell 2017; Jaldell, Lebnak, and Department of Transportation, who have provided invalu-
Amornpetchsathaporn 2014). Significant attention in the last able domain expertise to us.
several decades has been devoted to studying emergency in-
cidents and response. The broader goal of designing ERM Challenges and Lessons
systems is to enable communities to deal with emergency
response in a manner that is principled, proactive, and effi- While ERM forms a crucial component of the cities and gov-
cient. ERM systems based on data-driven models can help ernments, designing and deploying principled approaches
reduce both human and financial losses. Insights from prin- to ERM is challenging. We consider the following to be
cipled approaches can also be used to improve policies and the main challenges in the design and deployment of ERM
safety measures. Although such models are increasingly be- pipelines.
ing adopted by government agencies, emergency incidents 1. How to forecast incident occurrence? It has been noted
still cause thousands of deaths and injuries and also result that emergency incidents are generally difficult to pre-
in losses worth more than billions of dollars directly or indi- dict due to the inherent random nature of such incidents
rectly each year (Hattis 2015). and spatially varying factors (Shankar, Mannering, and
ERM can be divided into five major components: 1) mit- Barfield 1995; Mukhopadhyay et al. 2017). Incidents are
igation, 2) preparedness, 3) detection, 4) response, and 5) highly sporadic, making it particularly difficult to learn
recovery. The stages of ERM systems are heavily inter- models of incident occurrence. For example, well-known
linked (Mukhopadhyay et al. 2020). Mitigation involves sus- regression models such as Poisson regression and neg-
AAAI Fall 2020 Symposium on AI for Social Good. ative binomial regression have been shown to perform
Copyright c 2020 for this paper by its authors. Use permitted un- poorly on accident data due to the prevalence of zero
der Creative Commons License Attribution 4.0 International (CC counts (Mukhopadhyay et al. 2020). Incident data has also
BY 4.0). been shown to be particularly sensitive to the scale of
spatial and temporal resolutions, thereby making it dif- under which response takes place is dynamic. Consider a
ficult to perform meaningful inference. There are several decision-theoretic model for dispatching responders. For
approaches that have shown to alleviate these concerns. example, see the semi-Markov decision process formu-
First, identifying clusters of incidents (both spatial and lation to minimize expected response times to accidents
non-spatial) has shown to balance model variance and (Mukhopadhyay, Wang, and Vorobeychik 2018). An ap-
spatial heterogeneity particularly well (Sasidharan, Wu, proach to solve such large-scale decision problems is to
and Menendez 2015; Mukhopadhyay et al. 2017). Sec- use a simulator to find a policy that picks the optimal ac-
ond, dual state models like zero-inflated Poisson models tion given the state of the world. However, in urban areas,
can be used to address the issue of a high number of zero events like road closures, constructions, or increased traf-
counts in data (Qin, Ivan, and Ravishanker 2004). fic due to a public gathering can drastically alter the distri-
2. When to optimize? Arguably, the most important com- bution of incidents. Further, ambulances can be unavail-
ponent of ERM pipelines is to dispatch responders when able due to breakdowns or maintenance. In such cases,
incidents occur. While resource allocation and dispatch it is crucial that the actual state of the environment is
to emergency incidents evolve in highly uncertain and taken into account while creating allocation and dispatch
dynamic environments, the expectation is that response decisions. One approach is to create high-fidelity mod-
is very timely (Felder and Brinkmann 2002; Mukhopad- els of covariates such as traffic and weather. Such mod-
hyay, Wang, and Vorobeychik 2018). Approaches to op- els can then be used in online forecasting models, that
timize dispatch typically focus on decision-making after can accommodate incoming streams of updated infor-
an incident occurs (Toro-Dı́Az et al. 2013; Mukhopad- mation (Mukhopadhyay et al. 2019). Similarly, decision-
hyay, Wang, and Vorobeychik 2018; Keneally, Robbins, theoretic approaches that can quickly compute promising
and Lunday 2016). However, our conversations and col- actions for the current state of the world can be more
laborations with first responders revealed that there is lim- valuable in emergency response than approaches that find
ited applicability of such approaches in practice due to policies for the entire state space of the problem (Pettet
two important reasons. First, response to incidents occurs et al. 2020).
almost instantaneously after a report is received. Although The principled design of ERM systems is an important
optimizing dispatch can minimize response times in the problem faced by communities. As smart and connected
long run, time spent to optimize dispatch after incidents communities evolve, they present both opportunities and
occur is perceived as costly in the field. Second, it is al- challenges to manage ERM systems. In this short paper,
most impossible to judge the severity of an incident from we highlight common problems and lessons learned through
a call for service. Consequently, it is imperative for first our experience in designing ERM systems over the last few
responders to follow a greedy strategy and dispatch the years.
closest available responder to incidents. An alternative
approach is to periodically optimize the spatial distribu- Acknowledgement
tion of responders between incidents. Pettet et al. (Pettet
We would like to acknowledge the National Science Foun-
et al. 2020) introduced this idea recently for emergency
dation and the Center of Automotive Research at Stanford
response. While there are challenges with respect to scal-
for funding this research. We would also like to thank the
ability of such an approach, planning between incidents
Nashville Fire Department (NFD) and the Tennessee De-
is much more applicable in the field, as it does not violate
partment of Transportation (TDOT) for collaborating with
constraints under which first responder operate.
us and providing invaluable knowledge about the intricacies
3. How to model communication? Approaches to tackle of emergency response.
emergency response typically assume that the agents can
observe the world completely and communicate with cen- References
tralized servers and each other. This assumption is usually Felder, S.; and Brinkmann, H. 2002. Spatial allocation of emer-
satisfied in practice. However, in scenarios that involve gency medical services: minimising the death rate or providing
natural disasters (like floods, earthquakes, wildfires, etc.) equal access? Regional Science and Urban Economics 32(1): 27–
communication mechanisms can break down and power 45.
failures are common. In such scenarios, it is important that Hattis, S. H. 2015. Crime in the United States. Lanham Bernan
agents can optimize response based on information gath- Press.
ered locally. One way to approach such a problem is to
Jaldell, H. 2017. How important is the time factor? Saving lives
design distributed approaches for ERM, in which agents using fire and rescue services. Fire Technology 53(2): 695–708.
can optimize their own decisions (Pettet et al. 2020). This
is feasible since modern agents (ambulances, for example) Jaldell, H.; Lebnak, P.; and Amornpetchsathaporn, A. 2014. Time
is money, but how much? The monetary value of response time for
are equipped with laptops. While distributed approaches
Thai ambulance emergency services. Value in Health 17(5): 555–
perform worse than their centralized counterparts, they 560.
provide the benefit of performance in scenarios where
communication is challenging. Keneally, S. K.; Robbins, M. J.; and Lunday, B. J. 2016. A markov
decision process model for the optimal dispatch of military medical
4. How to model the environment? A problem in design- evacuation assets. Health Care Management Science 19(2): 111–
ing approaches to ERM is that environmental conditions 129.
Mukhopadhyay, A.; Pettet, G.; Samal, C.; Dubey, A.; and Vorob-
eychik, Y. 2019. An Online Decision-theoretic Pipeline for Re-
sponder Dispatch. In International Conference on Cyber-Physical
Systems (ICCPS), 185–196. ACM.
Mukhopadhyay, A.; Pettet, G.; Vazirizade, S.; Lu, D.; Baroud, H.;
Jaimes, A.; Vorobeychik, Y.; Kochenderfer, M.; and Dubey, A.
2020. A Review of Emergency Incident Prediction, Resource Al-
location and Dispatch Models.
Mukhopadhyay, A.; Vorobeychik, Y.; Dubey, A.; and Biswas, G.
2017. Prioritized Allocation of Emergency Responders based on a
Continuous-Time Incident Prediction Model. In International Con-
ference on Autonomous Agents and MultiAgent Systems (AAMAS),
168–177.
Mukhopadhyay, A.; Wang, Z.; and Vorobeychik, Y. 2018. A Deci-
sion Theoretic Framework for Emergency Responder Dispatch. In
International Conference on Autonomous Agents and MultiAgent
Systems (AAMAS), 588–596.
Pettet, G.; Mukhopadhyay, A.; Kochenderfer, M.; Vorobeychik, Y.;
and Dubey, A. 2020. On algorithmic decision procedures in emer-
gency response systems in smart and connected communities. In
International Conference on Autonomous Agents and Multiagent
Systems (AAMAS).
Qin, X.; Ivan, J. N.; and Ravishanker, N. 2004. Selecting exposure
measures in crash rate prediction for two-lane highway segments.
Accident Analysis & Prevention 36(2): 183–191.
Sasidharan, L.; Wu, K.-F.; and Menendez, M. 2015. Exploring the
application of latent class cluster analysis for investigating pedes-
trian crash injury severities in Switzerland. Accident Analysis &
Prevention 85: 219–228.
Shankar, V.; Mannering, F.; and Barfield, W. 1995. Effect of road-
way geometrics and environmental factors on rural freeway acci-
dent frequencies. Accident Analysis & Prevention 27(3): 371–389.
Toro-Dı́Az, H.; Mayorga, M. E.; Chanta, S.; and Mclay, L. A. 2013.
Joint location and dispatching decisions for emergency medical
services. Computers & Industrial Engineering 64(4): 917–928.