Self-Attentional Credit Assignment for Transfer in Reinforcement Learning

Authors: Johan Ferret, Raphael Marinier, Matthieu Geist, Olivier Pietquin

IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we aim to answer the following questions: can SECRET improve the sample efficiency of learning for RL agents? Does it generalize and/or transfer? How does it compare to transfer baselines? Is the credit assigned by SECRET interpretable?
Researcher Affiliation Industry Johan Ferret , Rapha el Marinier , Matthieu Geist and Olivier Pietquin Google Research, Brain Team {jferret, raphaelm, mfgeist, pietquin}@google.com
Pseudocode No The paper describes the methodology in prose and mathematical formulations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about open-sourcing the code for the described methodology or provide a link to a code repository.
Open Datasets Yes We use the keys doors puzzle 3D environment from DMLab [Beattie et al., 2016]
Dataset Splits No The paper discusses training on a 'source distribution' and evaluating on 'target environments' or 'held-out environments' but does not provide specific numerical train/validation/test splits or detailed splitting methodologies.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments.
Software Dependencies No The paper mentions general algorithms and architectures such as 'Q-learning', 'PPO', 'Transformer decoder', and 'convolutional layers', but does not provide specific version numbers for any software dependencies or libraries.
Experiment Setup Yes We use Q-learning [Watkins and Dayan, 1992] (tabular, with a learning rate of 0.1 and ϵ = 0.1) for experiments in Triggers except for out-of-domain transfer to environments with modified dynamics where we use DQN [Mnih et al., 2015]. We use PPO [Schulman et al., 2017] for in-domain experiments in DMLab, with identical hyperparameters as in Episodic Curiosity [Savinov et al., 2019], whose code is open-source... In Triggers experiments, we use 128 units per dense layer, 32 convolutional filters and a single convolutional layer to process partial states. We use a dropout rate of 0.1 after dense layers, a dropout rate of 0.2 in the self-attention mechanism and in the normalization blocks of the Transformer. Class weights in the loss function are set to w(1) = w( 1) = 0.499, w(0) = 0.002. In DMLab experiments, we use 16 convolutional filters and two convolutional layers to process partial states, and otherwise identical hyperparameters.