Policy Gradient for Rectangular Robust Markov Decision Processes

Authors: Navdeep Kumar, Esther Derman, Matthieu Geist, Kfir Y. Levy, Shie Mannor

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our RPG speeds up state-of-the-art robust PG updates by 2 orders of magnitude. and In the following experiments, we randomly generate nominal models for arbitrary state-action space sizes. Each experiment was averaged over 100 runs.
Researcher Affiliation Collaboration Navdeep Kumar Technion Esther Derman MILA, Université de Montréal Matthieu Geist Goodle Deepmind Kfir Levy Technion Shie Mannor Technion, NVIDIA Research
Pseudocode Yes Algorithm 1 RPG
Open Source Code Yes All codes and results are available at https://github.com/navdtech/rpg.
Open Datasets No In the following experiments, we randomly generate nominal models for arbitrary state-action space sizes.
Dataset Splits No The paper does not provide specific dataset split information (train/validation/test) as it uses randomly generated models rather than a pre-existing dataset with defined splits.
Hardware Specification Yes Hardware Experiments are done on the machine with the following configuration: Intel(R) Core(TM) i7-6700 CPU @3.40GHZ, size:3598MHz, capacity 4GHz, width 64 bits, memory size 64 Gi B.
Software Dependencies No All the experiments were done in Python using numpy, matplotlib.
Experiment Setup Yes Discount factor γ = 0.9, reward noise radius αs,a, αs = 0.1, transition noise kernel βs,a, βs = 0.01