Learning Causal Dynamics Models in Object-Oriented Environments

Authors: Zhongwei Yu, Jingqing Ruan, Dengpeng Xing

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on large-scale tasks indicate that OOCDM outperforms existing CDMs in terms of causal discovery, prediction accuracy, generalization, and computational efficiency.
Researcher Affiliation Academia 1Institute of Automation, Chinese Academy of Sciences, Beijing, P. R. China 2School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, P. R. China.
Pseudocode Yes Algorithm 1 Object-oriented causal discovery (Appendix E.3), Algorithm 2 Learning Object-oriented Causal Dynamics Model (Appendix E.4), and Algorithm 3 Planning with Cross Entropy Method (Appendix E.5) are provided.
Open Source Code Yes The source code of this work is made available at https://github.com/Ease Onway/oocdm.
Open Datasets Yes The Collect-Mineral-Shards (CMS) and Defeat-Zerglings-Baineling (DZB) environments are Star Craft II mini-games (Vinyals et al., 2017). The environment is adapted from the Walker2D environment, which is based on Mujoco, a popular physics simulator in the Open AI Gym platform (Brockman et al., 2016).
Dataset Splits No The paper discusses "i.d. (in-distribution) test data" and "o.o.d. (out-of-distribution) test data" for evaluation, but there is no explicit mention of a separate "validation" dataset split.
Hardware Specification Yes All models are trained and evaluated using one GPU (NVIDIA TITAN XP).
Software Dependencies No The paper mentions using "Python and Py Torch" but does not provide specific version numbers for these software components.
Experiment Setup Yes Table 14. The main hyper-parameters used in our experiments. This table lists various specific hyperparameters like de, dk, dv, dh, ε, nplan, H, α, β, λ, γ, k, k', niter, and nbatch for different environments.