Towards Hyperparameter-free Policy Selection for Offline Reinforcement Learning

Authors: Siyuan Zhang, Nan Jiang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform empirical evaluation on Open AI Gym [Bro+16], Atari games [BNVB13], and Mujoco [TET12]. ... For each algorithm, we consider different neural architectures, learning rates, and learning steps as hyperparameters to produce multiple candidate policies (and value functions) for selection; see Table 1 in Appendix C for details.
Researcher Affiliation Academia Siyuan Zhang Computer Science University of Illinois at Urbana-Champaign siyuan3@illinois.edu Nan Jiang Computer Science University of Illinois at Urbana-Champaign nanjiang@illinois.edu
Pseudocode Yes Based on this novel observation, we propose to search for a grid of discretization errors in BVFT and pick the resolution that minimizes the loss (Eq.(2)); see pseudocode in Appendix A.
Open Source Code No The paper does not provide an explicit statement or link for the open-sourcing of their code for the described methodology.
Open Datasets Yes We use standard offline datasets when available (RLUnplugged [Gul+21] for Atari, and D4RL [FKNTL21] for Mu Jo Co)...
Dataset Splits No The paper mentions re-sampling a subset of the dataset for policy selection (usually of size 50,000) for its evaluation, but does not explicitly describe train/validation/test splits for model training or for the data used in their method's evaluation in a reproducible manner.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) needed to replicate the experiment.
Experiment Setup Yes For each algorithm, we consider different neural architectures, learning rates, and learning steps as hyperparameters to produce multiple candidate policies (and value functions) for selection; see Table 1 in Appendix C for details. ... we propose to search for a grid of discretization errors in BVFT and pick the resolution that minimizes the loss (Eq.(2)); see pseudocode in Appendix A. ... Strategy 1 (using BVFT-PE-Q) slightly outperforms Strategy 2, but comes with an additional hyperparameter λ; we tuned it on Hopper and use the same constant in all experiments.