System Identification with Time-Aware Neural Sequence Models

Authors: Thomas Demeester3757-3764

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This section presents experimental results on two input/output datasets. The main research questions we want to investigate are (i) what is the impact of unevenly sampled data on prediction results with standard RNNs vs. with their time-aware extension, (ii) how important is the use of state update schemes that avoid the unit root issue, (iii) is there any impact from applying the time-aware models with higher-order schemes, and (iv) how are the output predictions affected while applying such higher-order schemes based only on sampled inputs? After describing the selected datasets (Section 4.1), the model architecture, and the training and evaluation setup (Section 4.2), we describe the experiments and discuss the results (Section 4.3).
Researcher Affiliation Academia Thomas Demeester Internet Technology and Data Science Lab, Ghent University imec thomas.demeester@ugent.be
Pseudocode No The paper describes mathematical equations for the models but does not include any explicit pseudocode blocks or algorithms.
Open Source Code Yes The code required to run the presented experiments is publicly available1. 1https://github.com/tdmeeste/Time Aware RNN
Open Datasets Yes We have selected two datasets from STADIUS’s Identification Database Da ISy (De Moor et al. 1997), a well-known database for system identification. CSTR Dataset We use the Continuous Stirred Tank Reactor (CSTR) dataset2. 2ftp://ftp.esat.kuleuven.be/pub/SISTA/data/process industry/ cstr.dat.gz Winding Dataset We also use the data from a Test Setup of an Industrial Winding Process3 (Winding). 3ftp://ftp.esat.kuleuven.be/pub/SISTA/data/process industry/ winding.dat.gz
Dataset Splits Yes It consists of a sequence of in total 7,500 samples, of which we used the first 70% for training, the next 15% for validation, and the last 15% for testing.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models) used for running the experiments.
Software Dependencies No The paper mentions using the Adam optimizer but does not specify versions for any programming languages, libraries, or other software dependencies.
Experiment Setup Yes Table 2: Hyperparameters tuned over ranges b {64, 512}, k {5, 10, 20, 30, 40, 60, 80, 100, 150}, and λ {0.001, 0.003, 0.01}. minibatch size b 512 512 512 64 state size k 20 100 10 10 learning rate λ 0.001 0.001 0.003 0.01