Finite-Time Analysis of Decentralized Single-Timescale Actor-Critic

Authors: qijun luo, Xiao Li

TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we conduct experiments to show the superiority of our algorithm over the existing decentralized AC algorithms.
Researcher Affiliation Academia Qijun Luo EMAIL School of Science and Engineering Shenzhen Research Institute of Big Data (SRIBD) The Chinese University of Hong Kong, Shenzhen Shenzhen, China Xiao Li EMAIL School of Data Science Shenzhen Institute of Artificial Intelligence and Robotics for Society (AIRS) The Chinese University of Hong Kong, Shenzhen Shenzhen, China
Pseudocode Yes Algorithm 1: Decentralized single-timescale AC (reward estimator version) Algorithm 2: Decentralized single-timescale AC (noisy reward version) Algorithm 3: Decentralized single-timescale NAC
Open Source Code No The paper does not contain an explicit statement about releasing source code or a link to a code repository.
Open Datasets No We adopt the grounded communication environment proposed in (Mordatch & Abbeel, 2018). Our task consists of N agents and the corresponding N landmarks inhabited in a two-dimension world, where each agent can observe the relative position of other agents and landmarks.
Dataset Splits No The paper describes the experimental environment and task setup but does not specify any training/test/validation dataset splits. The experiments appear to be conducted within a simulated environment rather than on a pre-defined dataset with splits.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup Yes For "SDAC-re" and "SDAC-noi", we set αk = 0.01(k + 1) 0.5, βk = 0.1(k + 1) 0.5, ηk = 0.1(k + 1) 0.5, Kc = 5, σ = 0.5, Kr = 2. For "DLDAC", we fix Tc = 50, T c = 10, T = 5, Nc = 10, N = 100, σ = 0.1 3, which is adopted by their paper (see comparisons under different hyper-parameters in Appendix A). We set α = 0.01, β = 0.1 for "DLDAC" since we observe that larger step sizes will result in divergence. We set αk = 0.01(k + 1) 0.5, βk = 0.1(k + 1) 0.5, Kr = 2, σ = 0.5 and examine the consensus periods Kc of 1, 5, 10, and 20, respectively.