JAT is a multi-purpose transformer agent capable of mastering a wide variety of sequential decision-making tasks. It uses a single network for all tasks and has shown rudimentary capabilities in NLP and CV tasks. The JAT project has released expert RL agents, the JAT dataset, and the JAT model. Future work includes improving the dataset, using offline RL, and implementing a smarter multi-task sampling strategy.
Table of contents
Introduction Datasets & expert policies The expert policies The JAT dataset JAT agent architecture Experiments and results The surprising benefits of predicting observations Conclusions What's next? A request for research Links Citation IntroductionDatasets & expert policiesJAT agent architectureExperiments and resultsConclusionsWhat's next? A request for researchLinksCitationShare this post