Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Identifying Causal Structure in Dynamical Systems
Authors: Dominik Baumann, Friedrich Solowjow, Karl Henrik Johansson, Sebastian Trimpe
TMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on a robot arm demonstrate reliable causal identification from real-world data and enhanced generalization capabilities. We evaluate the framework on three systems. First, we identify the causal structure of one arm of the robot Apollo (Kappler et al., 2018) shown in figure 3 in section 6.1. Then, we demonstrate the causal identification of a simulated quadruple tank process (cf. figure 5) in section 6.2. Lastly, we discuss a synthetic linear toy example. |
| Researcher Affiliation | Academia | Dominik Baumann EMAIL Department of Information Technology Uppsala University Uppsala, Sweden Friedrich Solowjow EMAIL Institute for Data Science in Mechanical Engineering RWTH Aachen University Aachen, Germany Karl H. Johansson EMAIL Division of Decision and Control Systems and Digital Futures KTH Royal Institute of Technology Stockholm, Sweden Sebastian Trimpe EMAIL Institute for Data Science in Mechanical Engineering RWTH Aachen University Aachen, Germany |
| Pseudocode | Yes | Algorithm 1 Pseudocode of the proposed framework. |
| Open Source Code | Yes | Code for both simulation examples is available at https://github.com/baumanndominik/identifying_causal_structure. |
| Open Datasets | No | Experiments on a robot arm demonstrate reliable causal identification from real-world data and enhanced generalization capabilities. We evaluate the framework on three systems. First, we identify the causal structure of one arm of the robot Apollo (Kappler et al., 2018) shown in figure 3 in section 6.1. Then, we demonstrate the causal identification of a simulated quadruple tank process (cf. figure 5) in section 6.2. Lastly, we present a synthetic, linear example. The paper does not provide concrete access information (link, DOI, repository, formal citation with authors/year) for publicly available datasets used in its experiments. |
| Dataset Splits | No | To investigate the generalization capability, we compare predictions of the model ˆfinit obtained from the initial system identification and the model ˆfcaus that was learned after revealing the causal structure. We use the same training data to estimate the model parameters in both cases. However, for ˆfcaus, we leverage the obtained knowledge of the causal structure when estimating parameters. In contrast, for ˆfinit we do not take any prior knowledge into account. As test data, we use an experiment that was conducted to investigate the influence of the initial condition of joint 3 on the other joints and let both models predict the trajectory of joint 1. The paper describes using "training data" and "test data" but does not provide specific details on how these datasets were split (e.g., percentages, sample counts, or explicit splitting methodology). |
| Hardware Specification | No | We identify the causal structure of one arm of the robot Apollo (Kappler et al., 2018). The paper refers to the robot system being experimented on, but does not specify the computational hardware (e.g., GPU/CPU models, memory) used for running the experiments or training the models. |
| Software Dependencies | No | We use the SINDy implementation from de Silva et al. (2020), We use the implementation provided in de Silva et al. (2020), We use the PCMCI algorithm proposed in Runge et al. (2019b), We use the implementation and parameter settings provided in https://github.com/jakobrunge/tigramite. While the paper mentions several algorithms and their implementations, it does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | We design input trajectories of 100 time steps for each experiment, repeat the experiment ten times, and use collected data from all experiments for hypothesis testing. ...We use ν = 1... We discretize the quadruple tank system with a time-step of 100 ms. ... For the initial model learning, we excite the system for 5000 time steps. During excitation, the input is sampled from a univariate distribution with an interval [0, 60]. ... We choose ν = 10 in equation 9 to avoid false positives. |