GradientDICE: Rethinking Generalized Offline Estimation of Stationary Values

Authors: Shangtong Zhang, Bo Liu, Shimon Whiteson

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we provide empirical results demonstrating the advantages of Gradient DICE over Gen DICE and Dual DICE.
Researcher Affiliation Academia 1University of Oxford 2Auburn University. Correspondence to: Shangtong Zhang <shangtong.zhang@cs.ox.ac.uk>.
Pseudocode Yes Algorithm 1 Projected Gradient DICE
Open Source Code Yes The implementations are made publicly available for future research.4 4https://github.com/Shangtong Zhang/Deep RL
Open Datasets Yes We consider two variants of Boyan s Chain (Boyan, 1999) as shown in Figure 4.
Dataset Splits No No specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning was found.
Hardware Specification No The paper mentions 'NVIDIA' in the acknowledgments for an 'equipment grant', but does not specify the exact hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No No specific software dependencies with version numbers (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4) were mentioned.
Experiment Setup Yes We use neural networks to parameterize τ and f, each of which is represented by a two-hidden-layer network with 64 hidden units and Re LU (Nair & Hinton, 2010) activation function. ... We train each algorithm for 10^3 steps and examine MSE(ρ) .= 1/2(ργ(π) − ρ̂γ(π))^2 every 10 steps... We use SGD to train the neural networks with batch size 128. The learning rate α and the penalty coefficient λ are tuned from {0.01, 0.005, 0.001} and {0.1, 1} with grid search to minimize MVE(ρ) at the end of training.