A Multi-Agent Reinforcement Learning-Based Framework for Forecasting Terrorist Collaboration and Predicting Future Alliances

Terrorist activity has increased over the years, leading to the rise of new criminal organizations, the persistence of incidents, and increased collaboration and coordination among criminal entities. This study proposes a framework based on multi-agent reinforcement learning (MARL) to forecast terrorism collaboration dynamics from time-series data and predict future collaborations. Firstly, we retrieve data from the Global Terrorist Database for numerous countries and construct a terrorist collaboration network. Subsequently, we employ the cumulative time series data to construct cumulative temporal graphs, thereby facilitating the observation of the evolution of collaboration over time. Then, we design a reward function that quantifies the lethality of terrorist groups, the benefits of collaborations, the group’s role in the network and the effectiveness of the partnership. Finally, we use the learned parameters to generate unobserved terrorist collaboration networks and, therefore, to predict the future potential collaborations for terrorist groups. The research findings demonstrate that the MARL approach exhibits superior forecasting performance in predicting terrorist collaboration networks. Future research endeavours should explore the potential of AI in countering terrorist activities.