APS: Active Pretraining with Successor Features

Authors: Hao Liu, Pieter Abbeel

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

Reproducibility Variable Result LLM Response
Research Type Experimental When evaluated on the Atari 100k data-efficiency benchmark, our approach significantly outperforms previous methods combining unsupervised pretraining with task-specific finetuning.
Researcher Affiliation Academia 1University of California, Berkeley, CA, USA.
Pseudocode Yes Algorithm 1: Training APS
Open Source Code No No, the paper does not provide a link to open-source code or explicitly state that the code is available.
Open Datasets Yes We evaluate our approach on the Atari benchmark (Bellemare et al., 2013) where we apply APS to Dr Q (Kostrikov et al., 2020) and test its performance after fine-tuning for 100K supervised environment steps.
Dataset Splits No No, the paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) for training, validation, or test sets.
Hardware Specification No No, the paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No No, the paper mentions frameworks and prior works it builds upon (e.g., Dr Q, Adam optimizer, pycolab game engine) but does not provide specific version numbers for software dependencies.
Experiment Setup Yes We largely follow Hansen et al. (2020) for hyperparameters used in our Atari experiments, with the following three exceptions. [...] We use the Adam optimizer (Kingma & Ba, 2015) with an learning rate 0.0001. We use discount factor γ = .99. Standard batch size of 32. ψ is coupled with a target network (Mnih et al., 2015), with an update period of 100 updates.