Crime Mapping Using Artificial Intelligent Agents in Nairobi City County, Kenya

Mbani BO^1*, Odera PA² and Kenduiywo BK¹

¹Department of Geomatic Engineering and Geospatial Information System, Jomo Kenyatta University of Agriculture and Technology, Nairobi, Kenya

²Division of Geomatics, School of Architecture, Planning and Geomatics, University of Cape Town, South Africa

*Corresponding Author : Benson Onyango Mbani
Department of Geomatic Engineering and Geospatial Information System, Jomo Kenyatta University of Agriculture and Technology, Nairobi, Kenya
Tel: +0675870001
E-mail: mmbani@jkuat.ac.ke

Received: October 23, 2017 Accepted: November 04 2017 Published: November 11 2017

Citation: Mbani BO, Odera PA, Kenduiywo BK (2017) Crime mapping using Artificial Intelligent Agents in Nairobi City County, Kenya. Geoinfor Geostat: An Overview 5:4. doi: 10.4172/2327-4581.1000171

The analysis and prediction of crime remains quite a difficult task since the crime system is overly complex. In Nairobi, crime is increasing challenge to the police despite an increased effort targeting the vice. This is due to the fact that the underlying factors that lead to a surge in crime events, which include proliferation of
light firearms, presence of organized criminal outfits, inequitable resource distribution, poor urban land use planning policies, youth unemployment and substance abuse coupled with an ill prepared police force both in terms of resources and logistics, are still yet to be fully addressed. Consequently, the nexus between crime
occurrence and spatial location of the crime event is key relationship in modeling crime. This paper presents an agent-based spatialtemporal approach for modeling crime occurrences exploiting the convergence of the two technologies: artificial intelligence and Geographic Information Systems. Artificial Intelligence (AI) is
used to build human behavior into agents that explore the spatial environment autonomously while at the same time learning from experience. Specifically, the AI algorithm that is used in the study is a form of reinforcement learning referred to as q-learning. Reinforcement learning is a type of machine learning approach
which models, into the agents, the capability to find their ways along street networks, learning from experience in each iteration as the simulation proceeds. Three types of agents were designed in this simulation: Offender; Target and Guardian agents. The multi-agent simulation was developed in Netlogo software. Netlogo
environment allows a user to design an artificial environment comprising the three agents including possible crime occurrence locations after several iterations for generating crime patterns. In addition to designing agents that participate in a crime event, a risk terrain model was generated by overlaying a set of potential riskfactors that influence a crime event. These factors were first tested for their colocation with crime occurrence using a Chi square test and finally overlaid to generate a risk terrain surface. A validation to test the accuracy of the model was conducted by comparing the counts of crime as generated by the simulation against those reported by authorities. The metric used in the comparison is the spearman’s rank correlation coefficient. The validation yields a correlation coefficient of 0.4 indicating there is some degree positive correlation since for a perfect positive correlation, we expect a correlation coefficient of 1. The correlation coefficient is a bit low due to various assumptions made in this simulation such as only allowing agents to move along a street network and limiting the type of crime to street robbery alone. Moreover, the set of potential risk factors considered in the generation of the risk surface is finite when in fact in real life, the factors that motivate crime are complex, multidimensional and almost infinite in number.