Metropolitan-scale Agent-based Modelling of Street Robbery using Reinforcement Learning: Quantitative Assessment of Infrastructure Intervention

Although agent-based modelling of crime has made great progress in the last decades, drawing concrete conclusions from the modelling results that can be directly applied to real-world environments has thus far remained challenging, mainly due to the fact that existing studies use scaled-down scenarios (due to computational constraints) and employ relatively simple behavioural models for perpetrators. In order to study different hypotheses of street robbery at the scale of a realistic urban agglomeration, a model has been developed that fully incorporates the mobility behaviour of a civilian population at high spatial (1m) and temporal (1s) resolution, co-simulated alongside a perpetrator agent population that is endowed with the ability to learn through experience how to travel across and roam within the urban landscape, resulting in a stochastic "common-knowledge" behaviour that mimics the intelligence that real perpetrators possess about their own environments. This works extends with pedestrian travel an existing, high-resolution, mobility simulation framework of private vehicle and public transport. The dynamic decision-making and re-routing of perpetrator agents, the environmental variables at link-level, the realistic vision of perpetrators, and the perpetrator movement and travel learning, which is modelled with a state-of-the-art deep reinforcement learning algorithm, are integrated into the extended simulation framework. In order to facilitate modelling at this large scale (that is, millions of agents in a spatial extent of 10’000s of sq. km), the data processing and agent routing is performed on the CPU (Central Processing Unit), while the movement, perception and decision-making of perpetrator agents is performed on GPU (Graphical Processing Unit) hardware. This work is demonstrated on a scenario developed for the City of Cape Town, South Africa. The results show that perpetrator agents effectively optimise for the specified reward signals. Four situational factors that influence the occurrence of street robbery are systematically investigated, that is, the relative target value of different civilians, the availability of potential guardians, the illumination due to sunlight and public lighting installations, and the area-level socio-economic deprivation. The good spatial and temporal comparisons of simulated robberies with geocoded robbery data verify the accuracy of the final hypothesis specification using the extended simulation framework. This hypothesis and framework are then used to assess the effectiveness of public lighting as a tool to reduce street robbery. As an outcome of this work, criminologists, urban planners and policymakers will be able to use this approach to (i) investigate crime hypotheses in different urban environments and contexts, thereby improving our understanding of how different factors contribute to street robbery, and (ii) investigate the effectiveness of different infrastructure intervention strategies in the medium to long term.