A Dynamical View of the Question of Why
Mehdi Fatemi, Sindhu Gowda
TL;DR
The paper develops a process-based framework for causation in dynamical systems, reframing causal questions as RL problems over diffusion/MDP dynamics. By introducing grit $\Gamma_B$ and reachability $\Lambda_B$ and proving two fundamental lemmas, it provides a principled way to quantify how events contribute to future outcomes and to decompose this influence across state and action components. The approach yields formal causation criteria (including sufficiency and necessity notions) that separate true causal effects from mere correlations, and it demonstrates practical applicability through Pong and a Type-1 Diabetes simulator without requiring explicit structural models. This work advances causal inference in time-evolving systems, enabling data-driven discovery of cause-effect relations in complex, high-dimensional processes with potential broad impact across engineering and biomedical domains.
Abstract
We address causal reasoning in multivariate time series data generated by stochastic processes. Existing approaches are largely restricted to static settings, ignoring the continuity and emission of variations across time. In contrast, we propose a learning paradigm that directly establishes causation between events in the course of time. We present two key lemmas to compute causal contributions and frame them as reinforcement learning problems. Our approach offers formal and computational tools for uncovering and quantifying causal relationships in diffusion processes, subsuming various important settings such as discrete-time Markov decision processes. Finally, in fairly intricate experiments and through sheer learning, our framework reveals and quantifies causal links, which otherwise seem inexplicable.