Research map / Probabilistic and scientific modeling / Causal discovery
Learning Causal Dynamics Models in Object-Oriented Environments
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.
Extends causal dynamics models to large object-oriented environments by sharing class-level causal graphs and predictor parameters across same-class objects, with a CMI-based causal discovery and learning method that handles varying object counts.
Problem
Causal dynamics models (CDMs) learned via causal discovery do not scale to large-scale environments with many objects, due to the combinatorial number of candidate causal edges and sample-efficiency limits.
Approach
Introduce the Object-Oriented Causal Graph, which encodes class-level local and global causalities shared across all instances of a class, and an attention-based shared field predictor per state field; causal discovery uses class-level conditional mutual information and learning uses an average instance log-likelihood adapted to varying object counts.
Key results
- OOCDM recovers the ground-truth causal graph with 100% accuracy in Block_5, Block_10 and Mouse, versus at most 97.7% for baselines in those settings.
- On causal discovery time in CMS, OOCDM takes 7.4 s versus 209.8 s for CDL and 5.5e4 s for GRADER.
- OOCDM achieves the best o.o.d. average instance log-likelihood in Block_5 (13.5 vs. negative values for most baselines) and Mouse (11.2), and the best planning return in Mouse (80.1) and CMS (3.4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OOCDM instead of the strongest listed baseline for causal graph discovery | Causal discovery models · Block₂, Block₅, Block₁₀, and Mouse | causal graph accuracy | Helped |
| Use OOCDM instead of the fastest listed baseline for causal discovery | Causal discovery models · Block₅, Block₁₀, Mouse, CMS, and DZB | causal discovery time | Helped |
| Use OOCDM instead of OOFULL, which uses the full OOCG (ablation) | OOCDM · Mouse o.o.d. | average instance log-likelihood: -1.5e9 ± 3.0e9 -> 11.2 ± 17.2 | Helped |
| Use OOCDM instead of OOFULL, which uses the full OOCG (ablation) | OOCDM · Mouse o.o.d. planning | average episodic return: 62.2 ± 8.7 -> 75.1 ± 17.5 | Helped |
| Use OOCDM instead of OOFULL, which uses the full OOCG (ablation) | OOCDM · DZB planning | average episodic return: 269.8 ± 21.5 -> 266.2 ± 11.4 | Hurt |
| Use OOCDM instead of OOFULL, which uses the full OOCG (ablation) | OOCDM · Mouse seen and unseen tasks | episodic return: 77.0 ± 24.6 -> 94.8 ± 29.7; 70.8 ± 22.4 -> 88.8 ± 34.8 | Helped |
Surprising findings
- OOFULL, the dense object-oriented variant, collapses on o.o.d. data (e.g., -4.6e4 in Block_2) despite its object-oriented representation, showing causal structure is needed for generalization.
- CDL-A has better causal discovery accuracy than CDL but generalizes worse out-of-distribution, attributed to greater regression capacity leading to overfitting.
Limitations
- Requires a given object-oriented MDP representation (class and field definitions) as prior knowledge; learning OOMDP representations is left to future work.
- Assumes causation and result symmetry (dynamics symmetric within classes); asymmetric dynamics need auxiliary attributes and only yield a minimal approximate OOCG.
- Assumes no unobserved confounders and all variables observable.
Experimental setup
Datasets: Block environment (synthetic) Mouse grid world (synthetic) CMS StarCraft II mini-game DZB StarCraft II mini-game
Benchmarks: Block_2 Block_5 Block_10 Mouse CMS DZB
Metrics: structural Hamming distance accuracy of causal graph computation time of causal discovery average instance log-likelihood episodic return with MPC and cross-entropy planning
Baselines: CDL CDL-A GRADER TICSA MLP GNN OOFULL
Builds on: Causal Dynamics Learning (CDL) Factored MDP Object-Oriented MDP GRADER TICSA
Closest papers by meaning
- Building Minimal and Reusable Causal State Abstractions for Reinforcement Learning (AAAI 2024 · similarity 0.88)
- Fine-Grained Causal Dynamics Learning with Quantization for Improving Robustness in Reinforcement Learning (ICML 2024 · similarity 0.87)
- Hierarchical Reinforcement Learning with Targeted Causal Interventions (ICML 2025 · similarity 0.86)
- Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains (ICML 2025 · similarity 0.86)
- Sim-to-Real Causal Transfer: A Metric Learning Approach to Causally-Aware Interaction Representations (CVPR 2025 · similarity 0.86)
- Tackling Non-Stationarity in Reinforcement Learning via Causal-Origin Representation (ICML 2024 · 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.