Zero-Shot Reinforcement Learning via Function Encoders

Authors: Tyler Ingebrand, Amy Zhang, Ufuk Topcu

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate state-of-the-art data efficiency, asymptotic performance, and training stability in three RL fields by augmenting basic RL algorithms with a function encoder task representation.
Researcher Affiliation Academia 1University of Texas at Austin. Correspondence to: Tyler Ingebrand <tyleringebrand@utexas.edu>.
Pseudocode Yes Algorithm 1 Function Encoder
Open Source Code Yes Code: https://github.com/tyler-ingebrand/Function Encoder RL
Open Datasets Yes The testing environment is a modified Half-Cheetah environment (Towers et al., 2023) where the segment lengths, friction coefficient, and control authority are randomized within a range each episode...
Dataset Splits No The paper mentions 'training dataset' and 'testing environment' but does not provide explicit information about validation splits (e.g., percentages or sample counts for training, validation, and test sets).
Hardware Specification Yes Hardware All experiments on performed on a 9th generation Intel i9 CPU and a Nvidia Geforce 2060 with 6 GB of memory.
Software Dependencies No The paper does not specify software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x). It mentions 'transformers' as a class of models but not specific library versions.
Experiment Setup Yes We use b = 100 basis functions for all experiments. See A.6 for an ablation on how the hyper-parameters affect performance.