Stabilizing Backpropagation Through Time to Learn Complex Physics

Authors: Patrick Schnell, Nils Thuerey

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on three control problems show that especially as we increase the complexity of each task, the unbalanced updates from the gradient can no longer provide the precise control signals necessary while our method still solves the tasks. Our code can be found at https://github.com/tum-pbs/Stable BPTT.
Researcher Affiliation Academia Patrick Schnell & Nils Thuerey School of Computation, Information and Technology Technical University of Munich Boltzmannstr. 3, 85748 Garching, Germany {patrick.schnell,nils.thuerey}@tum.de
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Our code can be found at https://github.com/tum-pbs/Stable BPTT.
Open Datasets No For the training and test data set, we generate 256 initial states, created by placing a group of four evaders randomly at a minimum distance around the target state and the two drivers farther away in the outer parts of the system. (Section 4.1)
Dataset Splits No The paper mentions 'training' and 'test' data sets but does not specify a separate 'validation' split or its size/percentage.
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU, CPU models, or specific cloud resources) used for the experiments.
Software Dependencies No The paper mentions using 'Adam as optimizer' but does not specify software dependencies with version numbers (e.g., 'Python 3.8', 'PyTorch 1.9').
Experiment Setup Yes We choose Adam as optimizer to process the four different backpropagation vectors with a learning rate of 0.001 and a batch size of 8. This set of hyperparameters performed well across our tests, but we include an extensive search over 792 training runs with variations of the optimizer, learning rate and batch size in the appendix. (Section 4) We train with 256 initial states... (Section 4.1). We train for 1000 epochs (Section B.2).