Toward Synergism in Macro Action EnsemblesDownload PDF

28 Sept 2020 (modified: 05 May 2023)ICLR 2021 Conference Withdrawn SubmissionReaders: Everyone
Keywords: Automated machine Learning, Reinforcement learning, Macro action ensemble
Abstract: Macro actions have been demonstrated to be beneficial for the learning processes of an agent. A variety of techniques have been developed to construct more effective macro actions. However, they usually fail to provide an approach for combining macro actions to form a synergistic macro action ensemble. A synergistic macro action ensemble performs better than individual macro actions within it. Motivated by the recent advances of neural architecture search, we formulate the construction of a synergistic macro action ensemble as a sequential decision problem and evaluate the ensemble in a task. The formulation of sequential decision problem enables coherency in the macro actions to be considered. Also, our evaluation procedure takes synergism into account since the synergism among the macro action ensemble exhibits when jointly used by an agent. The experimental results show that our framework is able to discover synergistic macro action ensembles. We further perform experiments to validate the synergism property among the macro actions in an ensemble.
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics
One-sentence Summary: Learning to construct a synergistic macro ensemble.
Reviewed Version (pdf): https://openreview.net/references/pdf?id=bCfGIrj6rZ
5 Replies

Loading