An Experimental Evaluation of Computational Techniques for Planning and Assessment of International Interventions

We describe the experimental methodology developed and employed in a series of experiments within the Defense Advanced Research Projects Agency (DARPA) Conflict Modeling, Planning, and Outcomes Exploration (COMPOEX) Program. The primary purpose of the effort was development of tools and methods for analysis, planning and predictive assessment of plans for complex operations where integrated political-military-economic-social-infrastructure and information (PMESII) considerations play decisive roles. As part of the program, our team executed several broad-based experiments, involving dozens of experts from several agencies simultaneously. The methodology evolved from one experiment to another because of the lessons learned. The paper presents the motivation, objectives, and structure of this interagency experiment series; the methods we explored in the experiments; and the results, lessons learned and recommendations for future efforts of such nature.

💡 Research Summary

The paper presents a comprehensive account of the experimental methodology employed in a series of inter‑agency experiments conducted under the Defense Advanced Research Projects Agency (DARPA) Conflict Modeling, Planning, and Outcomes Exploration (COMPOEX) program. The overarching goal of COMPOEX was to develop computational tools and analytical methods capable of supporting the planning, execution, and predictive assessment of complex international interventions where political, military, economic, social, infrastructure, and information (PMESII) dimensions interact in decisive ways.

The authors describe how a diverse cohort of experts—from the Department of Defense, State Department, other federal agencies, and academia—were brought together to jointly define the problem space, integrate data, generate alternative plans, and evaluate outcomes. The experimental design evolved across three major phases. In the first phase, participants established a common ontology for PMESII variables, standardized data formats, and agreed on a shared vocabulary. This groundwork was essential to avoid semantic mismatches that could corrupt subsequent modeling efforts.

In the second phase, the team constructed a hybrid simulation environment that combined system‑dynamics modeling of aggregate PMESII trends with agent‑based representations of individual actors and organizations. Causal links among variables were encoded in a network model, allowing the use of multi‑objective optimization algorithms—such as genetic algorithms and Pareto frontier exploration—to generate sets of candidate strategies that balance competing objectives (e.g., political legitimacy versus kinetic effectiveness) under resource constraints. Crucially, the optimization output was not presented as a single “best” plan; instead, trade‑off visualizations were provided so that human experts could apply judgment, contextual knowledge, and political feasibility considerations to select or refine strategies.

The third phase focused on predictive assessment. For each candidate strategy, thousands of Monte‑Carlo simulations were run to explore a wide range of possible future trajectories. Information operations, cyber influence, and adversary psychological responses were explicitly modeled, addressing a gap in traditional military‑centric simulations that often ignore the informational environment. Results were displayed on an interactive dashboard, enabling participants to observe emergent patterns, identify high‑risk scenarios, and iteratively adjust plans in a closed feedback loop.

Key findings emerged from the experiments. First, without prior consensus on data standards and terminology, inter‑agency collaboration quickly encounters modeling errors and communication breakdowns. Second, validation of the complex PMESII models required both quantitative comparison with historical operational data and qualitative assessment by domain experts; reliance on one source alone proved insufficient for establishing credibility. Third, algorithmically derived “optimal” solutions were not automatically executable; human expertise remained indispensable for assessing political acceptability, ethical constraints, and operational practicality. Fourth, incorporating the information domain dramatically enriched scenario diversity and yielded more realistic risk assessments, underscoring the importance of a full‑spectrum approach to intervention planning.

Based on these insights, the authors propose several recommendations for future efforts. They advocate for institutionalized mechanisms to maintain and evolve shared ontologies, continuous pipelines for feeding real‑world operational data into model validation, user‑centered design of human‑machine interfaces that emphasize transparency and explainability, and formal procedures for integrating model outputs into policy‑making cycles. Collectively, these guidelines aim to enhance the robustness, usability, and strategic relevance of computational tools for complex international interventions, paving the way for next‑generation decision‑support systems that can effectively blend automated analytics with expert judgment.