Learning from Active Human Involvement through Proxy Value Propagation

Authors: Zhenghao (Mark) Peng, Wenjie Mo, Chenda Duan, Quanyi Li, Bolei Zhou

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

Reproducibility Variable Result LLM Response
Research Type Experimental Human-in-the-loop experiments show the generality and efficiency of our method.
Researcher Affiliation Academia University of California, Los Angeles, University of Edinburgh
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Demo video and code are available at: https://metadriverse.github.io/pvp.
Open Datasets Yes We conduct experiments on various control tasks with different observation and action spaces. For continuous action space, we use three driving environments, Meta Drive safety benchmark [25], CARLA Town01 [8], and a customized driving environment built upon Grand Theft Auto V (GTA V), a popular video game. ... For discrete action space, we use Mini Grid Two Room task [4].
Dataset Splits Yes In Meta Drive, there exists a split of training and test environments, and we present the performance of the learned agent in a held-out test environment.
Hardware Specification Yes All experiments with humans are conducted on a local computer with an Nvidia Ge Force RTX 3080.
Software Dependencies No We implement most of the code with Stable-Baselines3 [37].
Experiment Setup Yes Hyper-parameters and other details are given in Appendix E and G. ... Table 7: PVP (Meta Drive) Hyper-parameter Value Discounted Factor γ 0.99 τ for Target Network Update 0.005 Learning Rate 0.0001 Steps before Learning Start 100 Steps per Iteration 1 Gradient Steps per Iteration 1 Train Batch Size 100 Q Value Bound 1