Crime Prediction and Prevention using Agent-Based Modeling

Transcription

1 Crime Prediction and Prevention using Agent-Based Modeling Computational Social Science Seminar PhD Student, Chair of Information Management Prof. Elgar Fleisch, D-MTEC Raquel Rosés Brüngger Seminar 17/10/2016 1

2 Problem description Seminar 17/10/2016 2

3 Crime prevention Crime has negative consequences Financial losses Reduction of emotional well-being But every crime is a source of information Aggregation of crimes reveal patterns Environmental factors influence crime Effective crime prevention efforts are information guided Crime prediction leads to more effective prevention Seminar 17/10/2016 3

4 Crime prediction techniques Analyze historical crime data Statistical techniques GIS (geographical information system) Miscellaneous techniques (data mining, machine learning, etc.) Drawbacks Fail to account for realistic representation of the environment Fail to account for changes in the environment Fail to account for actors and their interaction with environment Agent-Based Modeling (ABM) for crime prediction Seminar 17/10/2016 4

5 Previous work on ABM and crime Numerous ABM to test theories in criminology (e.g. Birks, Donkin, Wellsmith, 2008) Models including real data parameters and visual calibration e.g. (Liu, Wang, Eck, Liang, 2005) ABM for burglary including transportation network and statistical based human mobility patterns along the network (Peng, Kurland, 2014) ABM for burglary prediction including rather realistic representation of residential area and detailed offender architecture (e.g. Malleson, Evans, Jenkins, 2009; Malleson, Heppenstall, See, 2010; Malleson, See, Evans, Heppenstall, 2014) Seminar 17/10/2016 5

6 Method Seminar 17/10/2016 6

7 What is ABM? ABM is a powerful simulation technique, which replicates social phenomena for a better understanding of complex social systems Simplified simulation of reality Models dynamic processes and individual behavior Bares autonomous decision-making agents Rough example of ABM: Sims, SimCity. Seminar 17/10/2016 7

8 Characteristics of ABM Bottom up approach Distributed information system Autonomous agents, navigate environment System functioning is emergent (not hard coded) Seminar 17/10/2016 8

9 Intelligent agents Pursue complex and long-term goals Memory Learn form past experiences Improve skills over time Computational power intensive Mimic human-like behavior Seminar 17/10/2016 9

10 ABM to predict crime Seminar 17/10/

11 Crime prediction model architecture Figure 2: Crime prediction model architecture Seminar 17/10/

12 Research focus Automated calibration process Realistic environment (generators and attractors) Realistic offender s behavior Human dynamics and social media Crime data as a model parameter Seminar 17/10/

13 Type of Agents Offender agents Intelligent agents Criminal propensity Potentially responsible for crimes Population agents Represent population demographics Simple architecture (work, home, leisure) Deterring or attracting effect on offender agent (situation dependent) Seminar 17/10/

14 Environment Road and public transportation network Points of interest (shopping centers, police station, night clubs, etc.) Time (days, public holidays, etc.) Human dynamics Seminar 17/10/

15 Prototype and expected outcome Seminar 17/10/

16 Prototype: 1 month NYC burglary data Offender agent per type of crime Goal: offend Rule based actions: - Move along road/public transportation network - Decide whether to offend depending on environmental characteristics Composition of the environment influencing offender decision - Road/ public transportation network - Basic POI type - Residential vs. commercial area Seminar 17/10/

17 Repast Symphony - Java based platform for ABM Seminar 17/10/

18 Repast Symphony - Java based platform for ABM Seminar 17/10/

19 Expected Results and Impact of Crime Simulation Data-driven Crime Prediction Tool Intelligent Decision Support System Crime Prevention Experimental platform for policy makers and urban planners Seminar 17/10/

20 Thank you! Questions? Seminar 17/10/

21 Extra: Agent architecture example Figure 1: Basic Architecture PECS-Agent (Urban, Schmidt, 2001) Seminar 17/10/

22 Extra: Iterative evaluation of the model Figure 3: Evaluating complex computer models (Berk, 2008) Seminar 17/10/

23 Extra: Evaluation criteria 1. Visual comparison between simulated crime points and actual crime data points on the map (Malleson, 2010) 2. Center and standard deviational ellipse to describe the central and directional tendency of crime patterns (Peng & Kurland, 2014) 3. Local Moran s I to identify hotspots at a significant statistical level (Peng & Kurland, 2014). 4. Kernel density estimation method (KDE) to build density surface map of hotspots (Malleson, 2010) Seminar 17/10/