Continuous Doubly Constrained Batch Reinforcement Learning

Authors: Rasool Fakoor, Jonas W. Mueller, Kavosh Asadi, Pratik Chaudhari, Alexander J. Smola

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Over a comprehensive set of 32 continuous-action batch RL benchmarks, our approach compares favorably to state-of-the-art methods, regardless of how the offline data were collected.In this section, we evaluate our CDC algorithm against existing methods on 32 tasks from the D4RL benchmark [14]. We also investigate the utility of individual CDC regularizers through ablation analyses, and demonstrate the broader applicability of our extra-overestimation penalty to off-policy evaluation in addition to batch RL.
Researcher Affiliation Collaboration 1Amazon Web Services, 2University of Pennsylvania
Pseudocode Yes Algorithm 1 Continuous Doubly Constrained Batch RL
Open Source Code No The paper does not provide a concrete access link or explicit statement about the availability of the source code for the methodology described in this paper. It only mentions the code for a baseline method (CQL) from a reference: "The results for CQL are taken from the official author-provided codes [https://github.com/aviralkumar2907/CQL] of [29]."
Open Datasets Yes We compare CDC against existing batch RL methods... on 32 tasks from the D4RL benchmark [14]." Reference [14]: "J. Fu, A. Kumar, O. Nachum, G. Tucker, and S. Levine. D4rl: Datasets for deep data-driven reinforcement learning. ar Xiv:2004.07219, 2020."
Dataset Splits No The paper states: "Our training/evaluation setup exactly follows existing work [14, 17, 28, 29]," but it does not provide specific details on the training, validation, and test dataset splits (e.g., percentages, sample counts, or explicit citations to predefined splits) within its own text.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes In CDC, we can simply utilize the same moderately conservative value of ν = 0.75 used by [17]... CDC is able to achieve strong performance with a small ensemble of M = 4 Q-networks (used throughout this work)... Throughout, we use η = 0 & λ = 0 to refer to this baseline framework (without our proposed penalties)...