Causality and Complex Dynamic Systems: Bridging Statistical Inference and Control Theory

Statistical and Machine Learning methods for sequential decision making developed independently, creating parallel literatures that address mathematically identical problems under different terminologies. Recent work has indicated deep synergy between the two fields: Bruns-Smith and Pearl (2024) demonstrate that online reinforcement learning inherently captures causal relationships, and Bleile (2026), shows that advantage functions and mean-centered blip functions are mathematically equivalent objects under uniform policies. This panel examines how the relationship between reinforcement learning and causal inference transforms our approach to complex dynamic systems across multiple domains. One challenge in the control of complex dynamic systems is that causal inference in these systems is subject to limited completeness and contextual causality, particularly in biological and physiological domains where feedback loops and multi-scaling create temporal dependencies that static causal diagrams cannot capture (Diaz Ochoa, 2025). On the other hand, statistical methods, like synergistic, unique and redundant components (SURD), try to solve inherent problems of causality in complex dynamical systems by determining the specific nature of causal relationships, such as whether two variables are synergistic, i.e. one variable only influences another if it is paired with a second variable, or redundant. The synergy between causal AI and control theory has also been observed in practical applications such as medicine (Senn, 2013) and economics (Igami 2017). The panel addresses several fundamental questions that emerge at this intersection. First, when, if ever, do control objectives permit bias that inference objectives cannot tolerate? Second, how do we handle systems where traditional Markovian assumptions fail? For example, complex biological systems exhibit persistent dependencies and emergent properties that violate standard assumptions in both causal inference and reinforcement learning. Third, what new methodological opportunities emerge from explicit recognition of these translational synergies? Molak (2023) demonstrates practical implementation challenges when bridging causal inference and control frameworks in Python, revealing that computational tools developed independently in statistics and machine learning often implement identical mathematical operations using incompatible interfaces. The panel will explore how recent advances in causal discovery using reinforcement learning, heterogeneous treatment effect estimation, and adaptive experimental design create opportunities for practitioners to leverage methods from both traditions. In parallel with these developments, the panel will briefly situate this work within emerging 'causal roadmaps' that organize how we move from scientific questions to interventions, estimands, and algorithms (Bareinboim (2025), Hernán and Robins (2016), van der Laan and Petersen, (2014)).Discussion will emphasize practical implications for researchers working across the causal inference and machine learning divide, particularly in domains requiring sequential decision making under uncertainty with observational data constraints. Applications span clinical decision support systems where treatment sequences must be optimized from observational data, adaptive trial designs that balance exploration and exploitation while maintaining statistical validity, ecological systems management where intervention.