Weighted Policy Constraints for Offline Reinforcement Learning

Authors: Zhiyong Peng, Changlin Han, Yadong Liu, Zongtan Zhou

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our algorithm outperforms existing state-of-the-art offline RL algorithms on the D4RL offline gym datasets.
Researcher Affiliation Academia College of Intelligence Science and Technology, National University of Defense Technology, Changsha, China
Pseudocode Yes Algorithm 1: Weighted Policy Constraints
Open Source Code Yes The source code is available at https://github.com/qsa-fox/wPC.
Open Datasets Yes D4RL (Fu et al. 2020) is one of the main evaluation environments for offline RL, which consists of a wide of tasks and diverse datasets.
Dataset Splits Yes We run 1 million steps for training, evaluate policy every 5 thousand steps, and report the average normalized returns of 10 evaluation episodes as the score.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU or CPU models.
Software Dependencies No The paper does not provide specific version numbers for software dependencies (e.g., Python, PyTorch, TensorFlow versions) used in the experiments.
Experiment Setup Yes The only hyper-parameter attached to standard online RL is the α, which regulates the constraint strength. We set α to 0.1 for medium-expert datasets and 2.5 for others. ... Other hyper-parameters for TD3 components are presented in Table 3, and the neural network architectures are presented in Table 4.