Dynamics-Aware Embeddings

Authors: William Whitney, Rajat Agarwal, Kyunghyun Cho, Abhinav Gupta

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we empirically investigate how the learned Dyn E representations reshape the problem of reinforcement learning.
Researcher Affiliation Collaboration 1Department of Computer Science, New York University 2Robotics Institute, Carnegie Mellon University 3Facebook AI Research
Pseudocode No The paper describes the model and learning objective with equations, but no explicitly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes all of the code for Dyn E is available on Git Hub at https://github.com/dyne-submission/dynamics-aware-embeddings.
Open Datasets No The paper uses a dataset of '100K steps' generated from a uniformly random policy in the MuJoCo simulator, but does not provide concrete access information (link, DOI, citation) for this dataset to be publicly available.
Dataset Splits No The paper uses a dataset of '100K steps' for training representations but does not explicitly detail any train/validation/test splits, percentages, or sample counts.
Hardware Specification No All our experiments used recent-model NVidia GPUs.
Software Dependencies No The paper mentions software like MuJoCo, OpenAI Gym, TD3, SAC, and PPO implementations, but does not provide specific version numbers for any of these components.
Experiment Setup Yes We set our hyperparameters γ = λ = 10 2 across all environments. We use fully-connected networks for the action encoder ea and the conditional state predictor f. Each function has two hidden layers of 400 units. ... We set the dimension of the state embedding zs to 100. We set β = γ = 1... We use cyclic KL annealing (Liu et al., 2019) to improve convergence over a wide range of settings.