DC Field | Value | Language |
---|---|---|
dc.contributor.author | Lee, Su Young | ko |
dc.contributor.author | Choi, Sungik | ko |
dc.contributor.author | Chung, Sae-Young | ko |
dc.date.accessioned | 2019-12-13T07:35:01Z | - |
dc.date.available | 2019-12-13T07:35:01Z | - |
dc.date.created | 2019-11-24 | - |
dc.date.issued | 2019-12-10 | - |
dc.identifier.citation | NeurIPS 2019 | - |
dc.identifier.uri | http://hdl.handle.net/10203/268944 | - |
dc.description.abstract | We propose Episodic Backward Update (EBU) – a novel deep reinforcement learning algorithm with a direct value propagation. In contrast to the conventional use of the experience replay with uniform random sampling, our agent samples a whole episode and successively propagates the value of a state to its previous states. Our computationally efficient recursive algorithm allows sparse and delayed rewards to propagate directly through all transitions of the sampled episode. We theoretically prove the convergence of the EBU method and experimentally demonstrate its performance in both deterministic and stochastic environments. Especially in 49 games of Atari 2600 domain, EBU achieves the same mean and median human normalized performance of DQN by using only 5% and 10% of samples, respectively. | - |
dc.language | English | - |
dc.publisher | The Neural Information Processing Systems Foundation | - |
dc.title | The Neural Information Processing Systems Foundation | - |
dc.type | Conference | - |
dc.type.rims | CONF | - |
dc.citation.publicationname | NeurIPS 2019 | - |
dc.identifier.conferencecountry | CN | - |
dc.identifier.conferencelocation | Vancouver Convention Centre | - |
dc.contributor.localauthor | Chung, Sae-Young | - |
dc.contributor.nonIdAuthor | Lee, Su Young | - |
dc.contributor.nonIdAuthor | Choi, Sungik | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.