Research map / Probabilistic and scientific modeling / Causal discovery
A Dynamical View of the Question of Why
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a process-based theory of causation for diffusion/MDP time series that defines causes via "grit" (minimum probability of an effect) and reachability, computed as RL value functions and decomposed into per-variable contributions, validated on Atari Pong and a T1D simulator.
Problem
Causal reasoning about why an effect event occurs in multivariate stochastic time series (diffusion processes and MDPs), without known causal graphs, system equations, or infeasible interventions.
Approach
Defines causation between events via the expected change of "grit" (minimum probability of a future event), shows grit and reachability equal optimal values of two constructed RL reward MDPs (Value Lemma), and decomposes expected grit change into per-state/action component contributions via Itô's lemma (Decomposition Lemma).
Key results
- On Atari Pong, the method pinpoints the paddle-movement frames (49 onward) as causes of losing a point and identifies frame 51 as a sufficient cause, consistent with step-by-step game play.
- In the T1D simulator, the method attributes hypoglycemia at t in [312,313] to a change in subcutaneous insulin SI1 at t in [180,181] and, in the multi-intake scenario, rejects the earlier change because grit decreases (violating C2) and identifies SI1 at t in [510,511] as the sole cause.
- The paper reports qualitative results only; no quantitative baseline comparison is made because existing methods are argued to be inapplicable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| change in ball position/movement during the final steps before the point is lost | proposed method on Atari Pong · Atari Pong | grit of losing a point | Helped |
| change in SI1 during t ∈ [180,181] after a single insulin intake | proposed method on T1DMS · T1DMS, single insulin intake; hypoglycemia | grit of hypoglycemia | Helped |
| change in SI1 during t ∈ [510,511] after multiple insulin and meal intakes | proposed method on T1DMS · T1DMS, multiple insulin and meal intakes; hypoglycemia | grit of hypoglycemia | Helped |
Surprising findings
- Existing SCM and counterfactual causal methods are not used as baselines since they cannot be applied to these settings.
Limitations
- Higher-order moments of stochasticity beyond second order are dismissed.
- Assumes full-information state.
- Learned value functions may violate the theoretical properties (unity, null) to varying degrees due to approximation error.
Experimental setup
Datasets: Atari 2600 Pong T1DMS simulator (simglucose)
Benchmarks: Atari Pong T1DMS diabetes simulator
Metrics: expected change of grit per-variable gradient contributions
Builds on: DQN Lewis counterfactual theory Salmon process theory Halpern-Pearl actual causation
Closest papers by meaning
- A Practical Approach to Causal Inference over Time (AAAI 2025 · similarity 0.88)
- Systems with Switching Causal Relations: A Meta-Causal Perspective (ICLR 2025 · similarity 0.88)
- Causal learning with the invariance principle (NeurIPS 2026 · similarity 0.87)
- Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.87)
- On the Identification of Temporal Causal Representation with Instantaneous Dependence (ICLR 2025 · similarity 0.86)
- Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.