Adaptive Skip Intervals: Temporal Abstraction for Recurrent Dynamical Models

Authors: Alexander Neitz, Giambattista Parascandolo, Stefan Bauer, Bernhard Schölkopf

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of our approach by introducing two environments for which our approach is expected to perform well. Code to reproduce our experiments is available at https://github.com/neitzal/adaptive-skip-intervals. The experiments are ablation studies of our method. We would like to investigate the efficacy of adaptive skip intervals and whether the exploration schedule is beneficial to obtain good results. For each of our two environments, we compare four methods: (a) The recurrent dynamics model with adaptive skip intervals as described in Section 3. (ASI) (b) The dynamics model with adaptive skip intervals, but without any exploration phase, i.e. µ = 0. (ASI w/o exploration) (c) The dynamics model without adaptive skip intervals such that it is forced to predict every step (fixed ( t = 1)). (d) The dynamics model without adaptive skip intervals such that it is forced to predict every second step (fixed ( t = 2)). In each experiment we train with a training set of 500 trajectories, and we report validation metrics evaluated on a validation set of 500 trajectories.
Researcher Affiliation Academia 1Max Planck Institute for Intelligent Systems 2Max Planck ETH Center for Learning Systems 3aneitz@tue.mpg.de
Pseudocode Yes Algorithm 1: Dynamical model learning with ASI
Open Source Code Yes Code to reproduce our experiments is available at https://github.com/neitzal/adaptive-skip-intervals.
Open Datasets No The paper introduces custom environments (
Dataset Splits Yes In each experiment we train with a training set of 500 trajectories, and we report validation metrics evaluated on a validation set of 500 trajectories.
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU, CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions general software components like "neural network" and "Adam" optimizer, but it does not specify any software names with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 10.x).
Experiment Setup Yes For our experiments, we use a neural network with seven convolutional layers as the dynamics model f. Architectural details, which are the same in all experiments, are described in the Appendix. Like (Weber et al., 2017), we train f using a pixel-wise binary cross entropy loss. Hyperpararameter settings such as the learning rates are determined for each method individually by using the set of parameters which led to the best result (highest maximum achieved accuracy on the validation set), out of 9 runs each. We use the same search ranges for all experiments and methods. The remaining hyperparameters, including search ranges, are provided in the Appendix. For instance, as a value for the horizon H in the ASI runs, our search yielded optimal results for values of around 20 in both experiments.