On Algorithmic Decision Procedures in Emergency Response Systems in Smart and Connected Communities

2020·arXiv

ABSTRACT

1 INTRODUCTION

Emergency response management (ERM) is a critical problem faced by communities across the globe. First responders must attend to many incidents dispersed across space and time using limited resources. ERM can be decomposed into the following sub-problems — forecasting, planning, and dispatching. Although these have been examined independently, planning and dispatch decisions are dependent on accurate incident forecasting. Therefore, it is imperative that principled approaches are designed to tackle all three sub-problems. However, it is fairly common for ERM systems to follow myopic and straight-forward decision policies. For decades, the most common dispatching approach was to send the closest available responder to the incident (in time or space), after which the responder would return to its base or be reassigned. Such methods do not necessarily minimize expected response times [15]. As cities grow, population density, traffic dynamics and the sheer frequency of incidents make such methods stale and inaccurate. We systematically investigate the nuances of algorithmic approaches to ERM and describe how principled decision-making can aid emergency response.

Naturally, algorithmic approaches to emergency response typically combine a data-driven forecasting model to predict incidents with a decision-making process that provides dispatch recommendations. Canonical approaches towards modeling the decision process involve using a Continuous-Time Markov Decision Process (CTMDP)[11] or a Semi-Markovian Process (SMDP)[16], which are solved through dynamic programming. While the SMDP model provides a more accurate representation of ERM dynamics, it does not scale well for dynamic urban environments[13]. The trade-off between optimality and computational time has also been investigated by the use of Monte-Carlo based methods[13].

Despite such algorithmic progress and attention in recent years from the AI community[4, 13, 16, 18, 19, 22], there are still issues that impede the adoption of principled algorithmic approaches. We argue that a major problem lies in the very focus of most algorithmic approaches. Most ERM systems seek to perform decision-making after incidents occur. While such approaches guarantee optimality in the long run (with respect to response times), they de-prioritize response to some incidents. Our conversations with first-responders[6] revealed two crucial insights about this problem: 1) it is almost impossible to gauge the severity of an incident from a call for assistance and de-prioritize immediate response in anticipation of higher future rewards, and 2) Computer-Aided Dispatch systems (CAD)[21] typically enable a human agent to dispatch a responder in the span of 5-10 seconds. These insights explain why the closest responder is usually dispatched to an incident; it is too risky to de-prioritize incidents of unknown severity.

We raise an important conceptual question about algorithmic approaches to emergency response - is it feasible to optimize over dispatch decisions once an incident has happened? In this paper we argue that the crucial, practical period of principled decision-making is between incidents. This avoids the potential consequences of explicitly choosing to de-prioritize response to an incident to achieve future gain, but accommodates the scope of principled decision-making. Most ERM systems do not exploit the scope of dynamically rebalancing the spatial distribution of responders according to the need of the hour. This problem is challenging since optimizing responder distribution and response as a multi-objective optimization problem is usually computationally infeasible. Indeed, even Monte-Carlo based methods have previously been used with a restricted action space (only responding to incidents) to achieve acceptable computational latency[13]. We address this challenge by proposing two efficient algorithmic approaches to optimize over the spatial distribution of responders dynamically.

The second set of problems that impedes the adoption of algorithmic decision-making in ERM is related to resilience and efficiency. Data processing and decision-making for algorithmic dispatching usually occur in a centralized manner (typically at a central data processing center), which is then communicated to responders. ERM, however, clearly evolves in a multi-agent setting, in which the agents have the capacity to perform independent computation (most modern ambulances are equipped with laptops). In an extremely time-critical setting, especially during communication breakdowns often caused by disasters, it is crucial that such computing abilities are used, and distributed and parallelized algorithmic frameworks are designed. Also, centralized decision-making systems treat all agents as part of a monolithic object or state. This is redundant, as agents often operate independently (for example, an ambulance in one part of the city is usually not affected by an incident in a completely different or distant part). In this paper, we argue that decentralized planning could identify and utilize structure in the problem and save vital computational time.

Contributions: We focus on two problems in this paper (1) designing an approach that can accommodate rebalancing of resources to ensure efficient response, and (2) designing the ability for an emergency response system to be equipped to deal with scenarios that require decentralized planning with very limited communication. To this end, we start by modeling the problem of optimal response as a Multi-Agent Semi-Markov Decision Process (M-SMDP)[2, 9]. Then, we describe a novel algorithmic approach based on Multi-Agent Monte-Carlo Tree Search (M-MCTS)[3] that facilitates parallelized planning to dynamically rebalance the spatial distribution of responders. Our approach utilizes the computation capacity of each individual agent to create a partially decentralized approach to planning. Finally, we evaluate our framework using real-world data from Nashville, TN. We find that these approaches

Table 1: Notation lookup table

maintain system fairness while decreasing the average and variance of incident response times when compared to the standard procedure.

Outline: Through the rest of the paper, we describe the overall problem of emergency response and explain the algorithmic framework. We begin by providing a brief background regarding how ERM pipeline can be modeled technically, and how theoretical approaches to solution work in such situations. Then, we describe our algorithmic framework in detail, and finally, evaluate our framework using incident and response data from Nashville, TN. Table 1 can be used as a reference for the symbols we use.

2 SYSTEM MODEL

Our goal is to develop an approach for emergency responder placement and incident response in a dynamic, continuous-time and stochastic environment. We begin with several assumptions on the problem structure and information provided a-priori. First, we assume that we are given a spatial map broken up into a finite collection of equally-sized cells G, and that we are given an exogenous spatial-temporal model of incident arrival in continuous time over this collection of cells (we describe one such model later). Second, we assume that for each spatial cell, the temporal distribution of incidents is homogeneous. Our third assumption is that emergency responders are allowed to be housed in a set of fixed and exogenously specified collection of depots D. Depots are essentially a subset of cells that responders can wait in, and are analogous to fire-stations in the real-world. Each depot has a fixed capacity C(d) of responders it can accommodate at a time. We assume that when an incident happens, a free responder (if available) is dispatched to the site of the incident. Once dispatched, the time to service consists of two parts: 1) time taken to travel to the scene of the incident, and 2) time taken to attend to the incident. If no free responders are available, then the incident enters a waiting queue.

2.1 Incident Arrival

An important component of a decision-theoretic framework to aid emergency response is the understanding of when and where incidents occur. While our algorithmic framework can work with any forecasting model, we briefly describe the one that we choose to use: a continuous-time forecasting model based on survival analysis. It has recently shown state-of-the-art performance in prediction performance for a variety of spatial-temporal incidents (crimes, traffic accidents etc.)[15, 17, 18]. Formally, the model represents a probability distribution over inter-arrival times between incidents, conditional on a set of features, and can be represented as

where is a probability distribution for a continuous random variable T representing the inter-arrival time, which typically depends on covariatesw via the function. The model parameters can be estimated by the principled procedure of Maximum Likelihood Estimation (MLE) [10].

2.2 Decision-Making Process

The evolution of incident arrival and emergency response occur in continuous-time, and can be cohesively represented as a SemiMarkov Decision Process (SMDP) [16]. An SMDP system can be described by the tuple is a finite state space, A is the set of actions, P is the state transition function with being the probability with which the process transitions from state i to state j when action a is taken, T denotes the temporal transition with t(i, j,a) representing a distribution over the time spent during the transition from state i to state j under action represents the reward function, and is the discount factor.

To adapt this formulation to a multiagent setting, we model the evolution of incidents and responders together in a MultiAgent SMDP (MSMDP)[20], which can be represented as the tuple , where is a finite collection of agents and denotes the agent. The action space of the is represented by , and represents the joint ac- tion space. We assume that the agents are cooperative and work to maximize the overall utility of the system. The components and P are defined as in a standard SMDP. T represents a termination scheme; note that since agents each take different actions that could take different times to complete, they may not all terminate at the same time. An overview of such schemes can be found in prior literature [20]. We focus on asynchronous termination, where actions for a particular agent are chosen as and when the agent completes it’s last assigned action. Next, we define the important components of the decision process in detail.

States: A state at time t is represented by which consists of a tuple is a collection of cell indices that are waiting to be serviced, ordered according to the relative times of incident occurrence. corresponds to information about the set of agents at time t with . Each entry is a set , where is the position of responder is the destination cell that it is traveling to (which can be its current position), and is used to encode its current status (busy or available), all observed at the state of our world at time t. For the sake of convenience, we abuse notation slightly and refer to an arbitrary state simply by s and use the notation to refer to multiple states. We point out that our model revolves around states with specific events that provide the scope of decision-making. Specifically, decisions need to be taken when incidents occur, when responders finish servicing and while rebalancing the distribution of responders. We also make the assumption that no two events can occur simultaneously in our world. In case such a scenario arises, since the world evolves in continuous time, we can add an arbitrarily small time interval to segregate the two events and create two separate states.

Actions: Actions in our world correspond to directing the responders to a valid cell to either respond to an incident or wait. Valid locations include cells with pending incidents or any depot that has capacity to accommodate additional responders. For a specific agent , valid actions for a specific state are denoted by (some actions are naturally invalid, for example, if an agent is at cell k in the current state, any action not originating from cell k is unavailable to the agent). Actions can be broadly divided into two categories - responding and rebalancing. Responding actions refer to an agent actually going to the scene of an incident to service it. But agents could also be directed to wait at certain depots based on the likelihood of future incidents in the proximity of the said depot. We refer to such actions as rebalancing. Finally, we reiterate that the joint valid action space of all the agents and a particular instantiation of it are defined by A and a respectively, and that of a specific agent

Transitions: Having described the evolution of our world, we now look at both the transition time between states, as well as the probability of observing a state, given the last state and action taken. We define the former first, denoting the time between two states and by the random variable . There are four random variables of interest in this context. We denote the time between incidents by the random variable , the time to service an incident by , the time taken for a balancing step as and the time taken for a responder to reach the scene of an incident by . We overload these notations for convenience later. Specifically, we model using a survival model described in section 2.1. We model the service times () by learning an exponential distribution from service times using historical emergency response data, and we model rebalancing time () simply by the time taken by an agent to move to the directed cell.

We refrain from focusing on the transition function P, as our algorithmic framework only needs a generative model of the world and not explicit estimates of state transition probabilities.

Rewards: Rewards in SMDP usually have two components: a lump sum instantaneous reward for taking actions, and a continuous time reward as the process evolves. Our system only involves the former, which we denote by , for taking action a in state s. We define the exact reward function in section 3.3.

2.3 Problem Definition

Given state s and an agent set , the problem is to determine an action recommendation set , that maximizes the expected reward. The ith entry in contains a valid action for the ith agent.

Sovling this problem directly is hard due to its intractable state space. Further, the state transition functions are unknown and difficult to model in closed form, which is typical of urban scenarios where incidents and responders are modeled cohesively [16]. Finally, we have to consider the following practical constraints and limitations.

• Temporal constraints — emergency response systems can afford minimum latency ( 5-10 seconds in practice).

• Capacity constraints — each depot has a fixed agent capacity.

• Uniform severity constraint — all incidents must be responded to ‘promptly’, without making a judgement about its severity based on a report or a call.

• Wear and Tear — The overall distance agents travel should be controlled to limit vehicle wear and tear.

• Limited Communication - ERM systems must be equipped to deal with disaster situations, where communication is limited.

The temporal and uniform severity constraints make it difficult to justify implementing dispatch policies other than greedy; in order to improve upon greedy dispatch, some ‘good’ myopic rewards must be sacrificed for an increase in expected future rewards. Since it is very hard to predict the severity of an incident pre-dispatch, the decision process cannot determine if this sacrifice is acceptable. Therefore, in this work we focus on inter-incident planning while maintaining greedy dispatch decisions when an incident is reported. This approach gives the decision-maker more flexibility, as it can proactively position resources rather than reacting to incidents. Our problem then becomes how to distribute responders between incidents such that the greedy dispatching rewards are maximized.

3 REBALANCING APPROACH TO ERM

3.1 Problem Complexity

Dynamic rebalancing’s flexibility comes with an increase in complexity. Consider an example city with responders (i.e. agents) and |D| locations where responders could be stationed (called depots) that each can hold one responder. When making a dispatch decision at the time of an incident, a decision maker has at most possible choices: which responder to dispatch. If instead it is re-assigning responders across depots, there are significantly more choices. For example, with 20 and |D| = 30, there are 20 dispatching choices per incident, but 7possible assignments. This will only increase if depots have higher responder capacities.

Approaching the problem from this perspective requires solutions that can cope with this large complexity. One possible approach is to directly solve the SMDP model. Although the state transition probabilities are unknown, one can estimate the transition function by embedding learning into policy iteration[16]. This approach is unsuitable for rebalancing, as it is too slow even for the dispatch problem. A centralized MCTS approach suffers from the same shortcoming, barely satisfies the computational latency constraints in case of the dispatch problem[14]. Instead, we seek to exploit meaningful heuristics to propose computationally feasible rebalancing strategies. We begin by presenting our first approach, which focuses on using historical frequencies of incident occurrence across cells to assign responders.

3.2 Multi-Server Queue Based Rebalancing

One way to address the complexity of rebalancing is by considering an informed heuristic. A natural heuristic for ERM rebalancing is incident rate — each depot can be assigned responders based on the total rate of incidents it serves. Ultimately, our goal is to find a rebalancing strategy that minimizes expected response times. As a result, we first estimate the response time given a specific assignment of responders. Such a scenario can be modeled as a multi-server M/M/c queue [8]. For a given cell and depot, the response time for an M/M/c queue can be represented as

where is the mean service time of responders, is the number of responders stationed at the depot, denotes the rate of incident occurrence at the concerned cell, and is server utilization. The standard M/M/c model above needs slight adjustment to account for the fact that incidents at a cell can potentially be serviced by any depot, which are located at different distances from . Therefore, we consider a multi-class queue formulation in which a cell’s incident rate is split among each depot. Since depots closer to a cell are more likely to service its incidents, we split ’s incident arrival rate such that the fraction of rate incurred by a depot is inversely proportional to the distance to

The following system of linear equations can be used to split the arrival rate of a cell among depots D.

where the variable is the fraction of arrival rate of cell is shared by depot denotes the distance between depot d and cell , and is the depot closest to . Equation 2a ensures that the split rates for each cell sum to its actual arrival rate , and equation 2b ensures that the weighted ’s are inversely proportional to the relative distances between the depots and the cell. For convenience, we refer to the entire set of split rates by

The split rates provide a foundation for a responder rebalancing approach, given a few considerations. First, we might not have enough responders to meet the total demand based on ondly, the problem of evaluating response times in the context of emergency response is different than the standard M/M/c queue formulation, since travel times are not memoryless, and must be modeled explicitly. To address these issues, we design a scoring mechanism for evaluating a specific allocation of responders to

depots for a given . We denote this score by , a responder allocation can be scored by summing each depot d’s expected response time based on the queuing model (calculated using equation 1) and the overall time taken by responders to complete the rebalancing:

where is an indicator function which set to 1 only if depot d has at least one responder, and the functions responseTime and travelTime are used to denote the expected response time of a depot and travel times needed by agents to respond to incidents. The goal of an assignment method is then to find a responder allocation that minimizes this heuristic score. To minimize the total score we employ an iterative greedy approach, shown in algorithm 1. Once the best depots are found, responders are assigned to them based on their current distance from the depots. The approach dramatically decreases the computational complexity of rebalancing compared to a brute force search. The complexity for solving the system of linear equations {2a, 2b} is , as there are at most depots that could have a resource allocated. The rates are split for each cell and new depot under consideration d during each iteration of the greedy search in algorithm 1, which is repeated times to place each responder. This gives the overall algorithm a complexity of . Taking the same example given above with 20 and |D| = 30 and assuming |G| = 900 (based on our geographic area of interest and patrol areas chosen by local emergency responders), the complexity is 1 times less than a brute force search. While this approach is not inherently decentralized, each agent can perform these computations and take actions themselves, requiring minimal coordination. While straightforward and tractable, there are a few potential downsides to this approach. First, this policy does not take into account the internal state of the system.

For example, a responder might be on its way to respond to an incident, thereby rendering it unavailable for rebalancing. Secondly, it assumes that historical rates of incident arrival can be used to optimize responder placement for the future, thereby not considering how future states of the system affect a particular rebalancing configuration. To address these issues, we propose a decentralized Monte-Carlo Tree Search algorithm.

3.3 Decentralized MCTS Approach

Monte-Carlo Tree Search (MCTS) is a simulation-based search algorithm that has been widely used in game playing scenarios.

MCTS based algorithms evaluate actions by sampling from a large number of possible scenarios. The evaluations are stored in a search tree, which is used to explore promising actions. Typically, exploration policy is dictated by a principled approach like UCT[12]. A standard MCTS-based approach is not suitable for our problem due to the sheer size of the state-space in consideration coupled with the low latency that ERM systems can afford. Instead, we focus on a decentralized multi-agent MCTS (MMCTS) approach explored by Claes et. al [3] 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. This dramatically reduces the search space: in our case, at each evaluation step of a Monte-Carlo based approach, using a decentralized multi-agent search reduces the total number

of choices from the number of permutations

to only the number of depots |D|. To realize MMCTS for an ERM domain, some extensions need

to be made to standard UCT [7]. While an agent is building its own tree, it must model other agents’ behavior. Since this estimation is required at every step of every simulation by each agent, finding a model that strikes a balance between computation time and accuracy of predicted actions is vital. There are also global constraints on the system which mandate

agents maintain a minimal degree of coordination. For example,

the number of resources assigned to a depot cannot be higher than its capacity. We take this into account by adding a filtering step to the decision process. Similar to Map-Reduce [5], each agent sends their evaluated actions to a central planner which makes the final decisions while satisfying global system constraints. Next, we describe the architecture of our decentralized MMCTS

based algorithm.

• Reward Structure: At the core of an MCTS approach is an evaluation function that measures the reward of taking an action in a given state. For a state s in the tree of agent , we design the reward of taking an action a in s as

where refers to the total accumulated reward at the parent of state s in the tree, is the discount factor for future rewards, and the time since the beginning of the planning horizon . The evaluation function is split into cases reflecting the separate incident dispatch and balancing steps in our solution approach.

In a dispatch step, the reward is updated with the discounted response time to the incident . In a balancing step, we update the reward by the average distance traveled by the agents (we denote the distance traveled by agent while balancing due to action a in s by is an exogenous parameter that balances the trade-off between response time and distance traveled for balancing, and is set by the user depending on their priorities. Distance is not included during dispatch actions, as we always send the closest agent.

• Agents must have an accurate yet computationally cheap model of other agents’ behavior; we explore two such possible policies — (1) a naive policy that other agents will not rebalance, remaining at their current depot (referred to as Static Agent Policy), and (2) an informed policy, which is in the form of the Queue Rebalancing Policy described in the section 3.2. These are used to select actions for the other agents when building agent ’s search tree, and are represented by the ActionSelection(available agents, state) function in line 5 of algorithm 4.

• Rollout: When working outside the MMCTS tree, i.e. rolling out a state, a fast heuristic is used to estimate the score of a given action. We use greedy dispatch without balancing as our heuristic.

• Action Filtering: The dispatching domain has several global constraints to adhere to, including ensuring that an incident is serviced if agents are available and that depots are not filled over capacity. To meet these constraints, we propose a filtering step be added to the MMCTS workflow, similar to Map-Reduce. Once each individual agent has scored and ranked each possible action, these are sent to a centralized filter that chooses the final actions for each agent to maximize utility without breaking any constraints. Another way global constraints affect the workflow is that the

set of valid actions for an agent when they build their search tree

may not be the same as the valid actions when it comes time for them to make a decision. For example, consider two agents moves to a station and fills it to capacity, then agent cannot move to that station. To address this, we have agents evaluate every action they could possibly take when expanding nodes in the tree, even if those actions would cause an invalid state. As the filter assigns actions to other agents, some of these actions can become valid.

4 INTEGRATION FRAMEWORK

To realize an online ERM decision support system requires a framework of interconnected processes. Our integration framework is

built on our prior modular ERM pipeline work[13]. It includes the following components:

• A traffic routing model to support routing requests. • A model of the environment and how it changes over time, which is used by the incident prediction model.

• A model of the spatio-temporal distribution of incidents. • A decision process that makes dispatching recommendations based on the current state of the environment, responder locations, and future incident distributions. This framework is a natural choice as it decouples the decision process (our focus in this work) from other components. As it was designed for the centralized, post-incident dispatching approach, we make necessary changes to adapt it to our needs. The underlying discrete event simulation was generalized to accept events other than incident occurrence, such as periodic balancing events. The decision process was also extended to handle distributed, multi-agent approaches. An overview of the extended framework can be seen in figure 1. In our experiments we use a Euclidean distance based router, and the incident prediction model outlined in section 2. Due to

Figure 1: Extended Decentralized ERM Framework Overview

the framework’s modularity, these components can be replaced without affecting the decision process.

Incident Prediction Model: While the broader approach of rebalancing the spatial distribution of responders is flexible enough to work with any modular incident forecasting model, we provide a brief evaluation of forecasting using survival analysis. To this end, we generate forecasts 4 hours into the future at intervals of every half an hour for the entire test set, and then repeat the procedure 5 times to reduce variance and increase our confidence in the forecasts. Finally, we create a heatmap (average of all forecasted rates in the test set) to visualize the performance of the model in comparison to actual incidents (see figure 2). The forecasting models captures the high and low density areas fairly accurately, as well as the spatial spread of the incidents.

4.1 Experimental Design

We perform our evaluation on data from Nashville, TN, a major metropolitan area of USA, with a population of approximately 700,000. The depot locations are based on actual ambulance stations obtained from the city. Traffic accident data was obtained from the Tennessee Department of Transportation, and includes the location and time of each incident. The incident prediction model was trained on 35858 incidents occurring between 1-1-2018 and 1-1-2019, and we evaluated the decision processes on 2728 incidents occurring in the month of January, 2019.

Experimental Configuration and Assumptions: We limit the capacity of each depot to 1 in our experiments. This is motivated by two factors — first, it encourages responders to be geographically spread out to respond quickly to incidents occurring in any region of the city, and it models the usage of ad-hoc stations by responders, which are often temporary parking spots. While the responder service times to incidents are assumed to be exponential in the real world, we set them to a constant for these experiments. This ensures that the experiments across different methods and parameters are directly comparable. If deployed, however, proper service time distributions should be learned and sampled from for each ERM system. We set the total number of responders to 26, which is the actual number of responders in Nashville. We split the geographic area into 900, 1x1 mile square cells. This choice was a consequence of the fact that a similar granularity of discretization is followed by local authorities. To smooth out model noise, each agent evaluates 5 sampled incident chains from the generative model and averages the scores for each action across the playouts. The standard UCB1 [1] algorithm is used to select the most promising node during MCTS iterations. Finally, we augment the queue based rebalancing policy by adding a radius of influence

Figure 2: Heatmaps comparing average incident rates for the forecasting model (left) with actual incidents in Nashville, TN (right)

(RoI) for each cell. Only depots within a cell’s RoI are considered when splitting its rate to encourage even agent distribution and reduce computation time.

5 RESULTS AND DISCUSSION

We now discuss the results of the experiments for the two policies.

5.1 Queue Based Rebalancing Policy

We first compare the queue based rebalancing policy described in section 3.2 to the baseline policy of no rebalancing. In these experiments rebalancing occurred every half hour, and the incident rates were average historical rates from the training data. We tested several values (in miles) for the depots’ RoI, and compared the distributions of response times (figure 3a) and the rebalancing distance traveled by each responder (figure 3b).

Our first observation is that increasing the RoI does not necessarily increase performance; there is an optimal zone around RoI=3, implying that encouraging responders to spread out is beneficial. We also see that while Q-3’s median and 1st quartile response times remained fairly consistent with the baseline, the upper quartiles are reduced. This decreases the response time’s mean and variance, making the system more fair to all incidents. We also observe that Q-2 and Q-3’s responders traveled less than 1 mile on average each balancing step.

5.2 MMCTS Rebalancing

To determine the potential of the MMCTS rebalancing approach, we first compare the two agent action models described in section 3.3 (Static Agent Policy and Queue Rebalancing Policy) using an oracle, which has complete information regarding future incidents (this assumption takes the errors of the prediction model out of comparison and enables us to observe the best results that we can obtain) . We present the results for the response time distributions in figure 3c and the average responder distance traveled per rebalancing step in figure 3d.

Our first observation is that the MMCTS approach has high potential. Using an oracle, it is able to significantly decrease the response time distribution compared to the queue based policy above. This is not surprising given that a standard MCTS algorithm given perfect information should perform well given adequate time, but it demonstrates that the MMCTS extensions of independent

Figure 3: a) The response time distributions for each queue rebalancing policy experiment. b) Distribution of average miles traveled by each responder at each balancing step in the queue rebalancing policy experiments. The baseline approach has no rebalancing, so it is excluded. c) The response time distributions for each MMCTS experiment using an oracle. d) Distributions of average miles traveled by each responder at each balancing step of the MMCTS experiments using an oracle. e) The response time distributions for each MMCTS parameter search experiment. f) Distributions of average miles traveled by each responder at each balancing step of the MMCTS parameter search experiment.

action evaluation for each agent and action filtering are valid. Secondly, we see that MR-1 (using a static agent policy) outperforms MR-2 (using the queue rebalancing policy). Last, we observe that responders traveled between 2 and 4 miles on average each during balancing step in these experiments, which is significantly higher than the queuing approach.

Next, we examine a more the practical approach using the incident prediction model based on survival analysis. Since the static agent policy performed better in the oracle experiments, we use it for these experiments. There are several hyper-parameters that can affect the performance of the algorithm, including (1) MCTS Iteration Limit (2) Rebalancing Period - the amount of time between rebalancing steps (3) Distance Weight in Reward Function - this represents the importance of distance traveled for rewards (4) Look-ahead Horizon for MCTS.

We vary these parameters to see their effect on the system (see table 2). We present the response time distributions of MMCTS using the incident model in figure 3e, and the average responder distance traveled per rebalancing step in figure 3f. We observe that different parameter choices lead to different performance characteristics. For example, we see that changing the distance weight has a large impact on the distance responders travel; users with tight budgets for responder movement and maintenance will want to pay close attention to this parameter. Comparing the queue based policy with MMCTS, we see that both improve the response time distributions compared to the baseline. MMCTS is more configurable, but is also more sensitive to poor hyper-parameter choices. With proper hyper-parameter choices, both fulfil the constraints discussed in section 2.3 by having quick dispatching decisions, allowing for limited communication, and allowing users to control for distance traveled (i.e. wear and tear).

6 CONCLUSION

Principled approaches to Emergency Response Management (ERM) decision making have been explored, but have failed to be implemented into real systems. We have identified that a key issue with these approaches is that they focus on post-incident decision making. We argue that due to fairness constraints, planning should occur between incidents. We define a decision theoretic model for such planning, and implement both a heuristic search using queuing theory and a Multi Agent Monte Carlo Tree Search planner. We find

Table 2: Outline of the experimental runs performed and their cor- responding hyper-parameter choices. (When not indicated, parameters are set to values of M-1, the MMCTS Baseline in the table.)

that these approaches maintain system fairness while decreasing the average response time to incidents.

While the focus of this work is in the ERM domain, there are important takeaways for general agent-based systems: (1) Planning performance is dependent on the quality of the underlying event prediction models. (2) It is imperative to understand the needs and constraints for a target domain when designing a planning approach for it to be accepted in practice. (3) The computational capacity of “agents” has evolved in recent decades, and should be used to create decentralized planning approaches. Given these takeaways, we will explore the applicability of this framework to other domains where planning occurs over a spatial-temporally evolving process.

Acknowledgement: This work is sponsored by the National Science Foundation ( CNS-1640624, IIS-1814958, IIS-1905558), the Tennessee Department of Transportation and the Center for Automotive Research at Stanford (CARS). We thank our partners from Metro Nashville Fire Department, and Metro Nashville Information Technology Services in this work. We would also like to thank Hendrik Baier (CWI) for insights and helpful discussions regarding the paper. We also thank Sayyed Mohsen Vazirizade (Vanderbilt) for his help in reviewing the manuscript and providing important feedback.

REFERENCES

[1] Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. 2002. Finite-time analysis of the multiarmed bandit problem. Machine learning 47, 2-3 (2002), 235–256.

[2] Craig Boutilier. 1996. Planning, learning and coordination in multiagent decision processes. In Proceedings of the 6th conference on Theoretical aspects of rationality and knowledge. Morgan Kaufmann Publishers Inc., 195–210.

[3] Daniel Claes, Frans Oliehoek, Hendrik Baier, and Karl Tuyls. 2017. Decentralised online planning for multi-robot warehouse commissioning. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 492–500.

[4] One Concern. 2017. Artificial Intelligence: A GameChanger for Emergency Response. Technical Report. One Concern.

[5] Jeffrey Dean and Sanjay Ghemawat. 2008. MapReduce: Simplified Data Processing on Large Clusters. Commun. ACM 51, 1 (Jan. 2008), 107–113. https://doi.org/10. 1145/1327452.1327492

[6] Nashville Fire Department. 2018. Private Communication. (2018).

[7] Johannes Fürnkranz and Tobias Scheffer. 2006. Machine Learning: ECML 2006: 17th European Conference on Machine Learning, Berlin, Germany, September 18-22, 2006, Proceedings. Vol. 4212. Springer Science & Business Media.

[8] Natarajan Gautam. 2012. Analysis of queues: methods and applications. CRC Press.

[9] Mohammad Ghavamzadeh and Sridhar Mahadevan. 2006. Learning to Cooperate using Hierarchical Reinforcement Learning. (2006).

[10] Shenyang Guo. 2010. Survival analysis. Oxford University Press.

[11] Sean K Keneally, Matthew J Robbins, and Brian J Lunday. 2016. A markov decision process model for the optimal dispatch of military medical evacuation assets. Health care management science 19, 2 (2016), 111–129.

[12] Levente Kocsis and Csaba Szepesvári. 2006. Bandit based monte-carlo planning. In European conference on machine learning. Springer, 282–293.

[13] Ayan Mukhopadhyay, Geoffrey Pettet, Chinmaya Samal, Abhishek Dubey, and Yevgeniy Vorobeychik. 2019. An online decision-theoretic pipeline for responder dispatch. In Proceedings of the 10th ACM/IEEE International Conference on CyberPhysical Systems. ACM, 185–196.

[14] Ayan* Mukhopadhyay, Geoffrey* Pettet, Chinmaya Samal, Abhishek Dubey, and Yevgeniy Vorobeychik. 2019. An Online Decision-Theoretic Pipeline for Responder Dispatch. In ACM/IEEE International Conference on Cyber-Physical

Systems. ACM, 12–pages.

[15] Ayan Mukhopadhyay, Yevgeniy Vorobeychik, Abhishek Dubey, and Gautam Biswas. 2017. Prioritized Allocation of Emergency Responders based on a Continuous-Time Incident Prediction Model. In International Conference on Autonomous Agents and MultiAgent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 168–177.

[16] Ayan Mukhopadhyay, Zilin Wang, and Yevgeniy Vorobeychik. 2018. A Decision Theoretic Framework for Emergency Responder Dispatch. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, AAMAS 2018, Stockholm, Sweden, July 10-15, 2018. 588–596. http://dl.acm.org/ citation.cfm?id=3237471

[17] Ayan Mukhopadhyay, Chao Zhang, Yevgeniy Vorobeychik, Milind Tambe, Ken- neth Pence, and Paul Speer. 2016. Optimal Allocation of Police Patrol Resources Using a Continuous-Time Crime Model. In Conference on Decision and Game Theory for Security.

[18] Geoffrey Pettet, Saideep Nannapaneni, Benjamin Stadnick, Abhishek Dubey, and Gautam Biswas. 2017. Incident analysis and prediction using clustering and Bayesian network. In 2017 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computed, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI). IEEE, IEEE, San Francisco, CA, USA, 1–8.

[19] H. Purohit, S. Nannapaneni, A. Dubey, P. Karuna, and G. Biswas. 2018. Structured Summarization of Social Web for Smart Emergency Services by Uncertain Concept Graph. In 2018 IEEE International Science of Smart City Operations and Platforms Engineering in Partnership with Global City Teams Challenge (SCOPEGCTC). 30–35. https://doi.org/10.1109/SCOPE-GCTC.2018.00012

[20] Khashayar Rohanimanesh and Sridhar Mahadevan. 2003. Learning to take concurrent actions. In Advances in neural information processing systems. 1651–1658.

[21] Wikipedia contributors. 2019. Computer-aided dispatch — Wikipedia, The Free Encyclopedia. (2019). https://en.wikipedia.org/w/index.php?title= Computer-aided_dispatch&oldid=916096608 [Online; accessed 20-October-2019].

[22] Yisong Yue, Lavanya Marla, and Ramayya Krishnan. 2012. An efficient simulation-based approach to ambulance fleet allocation and dynamic redeployment. In Twenty-Sixth AAAI Conference on Artificial Intelligence.

Designed for Accessibility and to further Open Science