Optimal Exploration for Model-Based RL in Nonlinear Systems

Authors: Andrew Wagenmaker, Guanya Shi, Kevin G. Jamieson

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conclude with experiments demonstrating the effectiveness of our method in realistic nonlinear robotic systems1.
Researcher Affiliation Academia Andrew Wagenmaker Paul G. Allen School of Computer Science & Engineering University of Washington Seattle, WA 98195 Guanya Shi Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213 Kevin Jamieson Paul G. Allen School of Computer Science & Engineering University of Washington Seattle, WA 98195
Pseudocode Yes Algorithm 1 Optimal Exploration in Nonlinear Systems (informal)
Open Source Code Yes Code: https://github.com/ajwagen/nonlinear_sysid_for_control
Open Datasets No The paper uses simulated systems (drone, car, and a 1-D system example) which are internally generated, not publicly available datasets or benchmarks with specified access information.
Dataset Splits No The paper conducts experiments on simulated systems using 'episodes' of interaction. It does not provide explicit training, validation, or test dataset splits as one would for a standard machine learning dataset.
Hardware Specification Yes All experiments were run on a machine with 56 Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz CPUs, and 64GB RAM.
Software Dependencies No All code was implemented in Py Torch. However, no specific version numbers for PyTorch or Python are provided.
Experiment Setup Yes For all examples the noise is distributed as wh N(0, 0.1 I). In all cases we set γ2 = 10H (where γ2 is a bound on Eπexp[PH h=1 u h uh]), and we therefore let Πexp denote the set of all policies satisfying Eπexp[PH h=1 u h uh] γ2.