Abstract Disrupting terrorist networks remains a critical challenge for counter-crime units due to their adaptive, decentralized, and covert nature. Existing approaches provide valuable structural insights but often rely on static representations and do not capture the sequential and adversarial dynamics underlying real-world intervention processes. In this work, we propose an explainable, game-theoretic multi-agent reinforcement learning (MARL) framework for simulating and analyzing adaptive terrorist network disruption. We formulate the problem as a partially observable, sequential decision-making process between two agents: an attacker , representing a terrorist organization seeking to expand its operational or ideological influence, and a defender , modeling law enforcement entities tasked with disrupting influence and neutralizing key actors. The framework incorporates domain-informed reward functions that capture structural, behavioral, and resilience-related properties of networks, enabling both agents to learn policies through repeated interaction. A central contribution of this work is the integration of explainability into the learning framework, allowing the generation and quantitative evaluation of interpretable rationales for node-level intervention decisions. Rather than modeling the full complexity of radicalization processes, the proposed approach provides a stylized but tractable simulation environment for studying strategic interaction under uncertainty. Empirical evaluations on multiple real-world-inspired extremist network structures show that (i) disruption effectiveness improves with increasing intervention budget but exhibits non-monotonic and network-dependent behavior, (ii) outcomes are strongly shaped by attacker–defender strategy interactions rather than individual strategies alone, and (iii) learned policies produce consistent and structured explanation patterns that reveal underlying network vulnerabilities. These findings demonstrate that explainable MARL can provide actionable insights into adaptive intervention strategies and serve as a decision-support tool for intelligence-led policing in complex networked environments.
Dogan et al. (Sat,) studied this question.