Learning Hybrid Dynamics Models with Simulator-Informed Latent States

Authors: Katharina Ensinger, Sebastian Ziesche, Sebastian Trimpe

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

Reproducibility Variable Result LLM Response
Research Type Experimental 6 Experiments In this section, we show that: (i) Our KKL-RNN achieves equal or higher accuracy than baselines, especially in the partially OVS case; (ii) We learn a plausible split in OVS and non-OVS components with our method; (iii) Our method can buffer missing simulator inputs; (iv) We can easily incorporate properties as a decaying non-OVS part; (v) The concept can also be leveraged in the pure learning-based scenario. Table 1: RMSEs for Systems i)-iii) (mean (std)) over 5 independent runs.
Researcher Affiliation Collaboration 1 Bosch Center for Artificial Intelligence, Renningen, Germany 2 Institute for Data Science in Mechanical Engineering, RWTH Aachen Univeristy
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statement or link indicating that the source code for the methodology is openly available.
Open Datasets Yes ii) Double-torsion pendulum: We consider the measurements and the corresponding numerical simulation from Lisowski et al. (2020) and use the first 150 steps for training. ... iii) Drill-string system: We train on measurements provided in Aarsnes and Shor (2017) Fig. 14 and the corresponding simulator.
Dataset Splits No The paper mentions training data and performing predictions on the full trajectory, but it does not specify explicit training/validation/test splits, percentages, or absolute sample counts for each split. It uses a 'warmup phase' but this is not a dataset split.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions using a gated recurrent unit (GRU) but does not provide specific version numbers for software, libraries, or solvers used in the experiments.
Experiment Setup Yes The first R steps ˆy0:R are used as a warmup phase for the non-OVS residuum to obtain appropriate latent states. We train our model by computing an N-step rollout z0:N, yv 0:N and y0:N via Eq. (9) and minimizing the loss ˆθ = arg min θ y0:N ˆy0:N 2 + s0:N ˆs0:N 2 +λ yv 0:N 2, (10) where 2 denotes the MSE. We introduce a regularization factor λ R that allows to balance the influence of the learning-based component as it is typical for hybrid models (Takeishi and Kalousis 2021; Yin et al. 2021). Here, we learn T θ with an MLP... The trainable observation models gθ, hθ and rθ are modeled as linear layers. Here, we model the non OVS part f v θ in Eq. (8) with a gated recurrent unit (GRU). In the experiments, we apply exponential damping by bounding the RNN observation model rθ with an appropriate activation function and multiplying it with exp( at).