GRAND-VISION: An Intelligent System for Optimized Deployment Scheduling of Law Enforcement Agents
Figure 1. Agents patrol their assigned patrol regions. When
an incident occurs, the nearest agent is assigned to attend,
with the aim of meeting the response time QoS requirements.
Project Overview
Law enforcement agencies in dense urban environments,
faced with a wide range of incidents to handle and limited
manpower, are turning to data-driven AI to inform their policing strategy. In this project,
we present a patrol scheduling system called GRAND-VISION: Ground Response Allocation
and Deployment - Visualization, Simulation, and Optimization. The system employs deep learning to generate incident
sets that are used to train a patrol schedule that can accommodate varying manpower, break times, manual pre-allocations,
and a variety of spatio-temporal demand features. The complexity of the scenario results in a system with real world
applicability, which we demonstrate through simulation on
historical data obtained from a large urban law enforcement
agency.
My work: the first neural network WCGAN with linear regression
In the beginning of the project, we experimented a generative adversarial network to generate the incidents.
The data we have was crime's occurence, locations, urgency, numbers of police attended and engagement time.
For locations, we extracted latitude and longitude features.
For occurence, we extracted thirteen boolean data: eleven public holidays of Singapore,
whether the day is weekend or weekday, and the day within week.
We further encoded the cyclical continous features of time (into cosin and sine), such as day of weeks, day of years and week of months.
For other features, we keep them as is.
We used WCGAN to predice desired features of a crime which are latitude, longitude,
crimes' occurence in minutes from midnight, urgency, engagement time and
number of police attended. We conditioned on the boolean data of occurence.
However, there is still one missing piece required in WCGAN which is the number of samples we want it to generate. Thus,
we employed Ridge regression method to predict the number of incidents. The data we used for training is the cyclical continous features.
Figure 2. WCGAN with Ridge Regression.
My work: the second neural network STResnet with VAE
We observe that the crimes are usually occur in some specific locations and frequently at a cyclic interval, yet
the tabular representation in the first neural network does not capture both information. Thus we decided to research a spatio-temporal model
that can learn these meaningful patterns. We engineered an appropriate resolution in space and time so that the image's intensities
exhibits locality and stationary so that a CNN architecture could learn well.
We used STResnet[1] as a baseline architecture.
Because the deployment scheduler, which performs the allocations of on-duty agents in our systems,
requires different scenarios of crimes, we added VAE network into STResnet to generate multiple events.
The final architecture is called Stochastic STResnet is shown in Figure 3.
