Reinforcement Learning with Prototypical Representations

Authors: Denis Yarats, Rob Fergus, Alessandro Lazaric, Lerrel Pinto

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we discuss empirical results on using Proto RL for learning visual representations. (Section 5) and Proto-RL significantly improves upon Random exploration and APT across all environments, while being better than Curiosity based exploration in 7/8 environments. (Section 5.2)
Researcher Affiliation Collaboration 1New York University 2Facebook AI Research.
Pseudocode Yes The pseudo-code for our framework is provided in Appendix C. (Section 4.1) and Algorithm 1: Proto-RL (Appendix C).
Open Source Code Yes We open-source our code at https: //github.com/denisyarats/proto.
Open Datasets Yes We use the Deep Mind Control Suite (Tassa et al., 2018), a challenging benchmark for image-based RL.
Dataset Splits No The paper describes environment interactions and evaluation episodes, but it does not specify explicit training/validation/test dataset splits like percentages or sample counts for a static dataset. In RL, data is dynamically generated.
Hardware Specification No The paper does not provide specific hardware details such as GPU models (e.g., NVIDIA V100), CPU models, or cloud computing instance types used for the experiments.
Software Dependencies Yes Proto-RL is trained using Adam (Kingma & Ba, 2014)... We use SAC implementation from Yarats & Kostrikov (2020).
Experiment Setup Yes Hyper-parameters Proto-RL is trained using Adam (Kingma & Ba, 2014) with learning rate 10 4 and mini-batch size of 512. The downstream exploration hyper-parameter is = 0.2 and the number of cluster candidates is set to T = 4. We use SAC implementation from Yarats & Kostrikov (2020).