Stop Regressing: Training Value Functions via Classification for Scalable Deep RL

Value functions are a central component of deep reinforcement learning (RL). These functions, parameterized by neural networks, are trained using a mean squared error regression objective to match bootstrapped target values. However, scaling value-based RL methods that use regression to large networ...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:arXiv.org 2024-03
Hauptverfasser: Farebrother, Jesse, Orbay, Jordi, Vuong, Quan, Adrien Ali Taïga, Chebotar, Yevgen, Xiao, Ted, Irpan, Alex, Levine, Sergey, Castro, Pablo Samuel, Faust, Aleksandra, Kumar, Aviral, Agarwal, Rishabh
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title arXiv.org
container_volume
creator Farebrother, Jesse
Orbay, Jordi
Vuong, Quan
Adrien Ali Taïga
Chebotar, Yevgen
Xiao, Ted
Irpan, Alex
Levine, Sergey
Castro, Pablo Samuel
Faust, Aleksandra
Kumar, Aviral
Agarwal, Rishabh
description Value functions are a central component of deep reinforcement learning (RL). These functions, parameterized by neural networks, are trained using a mean squared error regression objective to match bootstrapped target values. However, scaling value-based RL methods that use regression to large networks, such as high-capacity Transformers, has proven challenging. This difficulty is in stark contrast to supervised learning: by leveraging a cross-entropy classification loss, supervised methods have scaled reliably to massive networks. Observing this discrepancy, in this paper, we investigate whether the scalability of deep RL can also be improved simply by using classification in place of regression for training value functions. We demonstrate that value functions trained with categorical cross-entropy significantly improves performance and scalability in a variety of domains. These include: single-task RL on Atari 2600 games with SoftMoEs, multi-task RL on Atari with large-scale ResNets, robotic manipulation with Q-transformers, playing Chess without search, and a language-agent Wordle task with high-capacity Transformers, achieving state-of-the-art results on these domains. Through careful analysis, we show that the benefits of categorical cross-entropy primarily stem from its ability to mitigate issues inherent to value-based RL, such as noisy targets and non-stationarity. Overall, we argue that a simple shift to training value functions with categorical cross-entropy can yield substantial improvements in the scalability of deep RL at little-to-no cost.
format Article
fullrecord <record><control><sourceid>proquest</sourceid><recordid>TN_cdi_proquest_journals_2941443136</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2941443136</sourcerecordid><originalsourceid>FETCH-proquest_journals_29414431363</originalsourceid><addsrcrecordid>eNqNjMsKwjAURIMgWLT_cMF1oU3S-thWiws32uK2XEtaUkJSk8bvN4If4GqGOYdZkIgyliV7TumKxM6NaZrSYkfznEXkVs9mgrsYrHBO6uEIjUWpQ4MHKi-g8rqbpdEO3hKhVBi0Xnb43aA3FuoOFT6VgJMQ4em6IcselRPxL9dkW52b8pJM1ry8cHM7Gm91QC098IxzlrGC_Wd9AGxMPrE</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>2941443136</pqid></control><display><type>article</type><title>Stop Regressing: Training Value Functions via Classification for Scalable Deep RL</title><source>Free E- Journals</source><creator>Farebrother, Jesse ; Orbay, Jordi ; Vuong, Quan ; Adrien Ali Taïga ; Chebotar, Yevgen ; Xiao, Ted ; Irpan, Alex ; Levine, Sergey ; Castro, Pablo Samuel ; Faust, Aleksandra ; Kumar, Aviral ; Agarwal, Rishabh</creator><creatorcontrib>Farebrother, Jesse ; Orbay, Jordi ; Vuong, Quan ; Adrien Ali Taïga ; Chebotar, Yevgen ; Xiao, Ted ; Irpan, Alex ; Levine, Sergey ; Castro, Pablo Samuel ; Faust, Aleksandra ; Kumar, Aviral ; Agarwal, Rishabh</creatorcontrib><description>Value functions are a central component of deep reinforcement learning (RL). These functions, parameterized by neural networks, are trained using a mean squared error regression objective to match bootstrapped target values. However, scaling value-based RL methods that use regression to large networks, such as high-capacity Transformers, has proven challenging. This difficulty is in stark contrast to supervised learning: by leveraging a cross-entropy classification loss, supervised methods have scaled reliably to massive networks. Observing this discrepancy, in this paper, we investigate whether the scalability of deep RL can also be improved simply by using classification in place of regression for training value functions. We demonstrate that value functions trained with categorical cross-entropy significantly improves performance and scalability in a variety of domains. These include: single-task RL on Atari 2600 games with SoftMoEs, multi-task RL on Atari with large-scale ResNets, robotic manipulation with Q-transformers, playing Chess without search, and a language-agent Wordle task with high-capacity Transformers, achieving state-of-the-art results on these domains. Through careful analysis, we show that the benefits of categorical cross-entropy primarily stem from its ability to mitigate issues inherent to value-based RL, such as noisy targets and non-stationarity. Overall, we argue that a simple shift to training value functions with categorical cross-entropy can yield substantial improvements in the scalability of deep RL at little-to-no cost.</description><identifier>EISSN: 2331-8422</identifier><language>eng</language><publisher>Ithaca: Cornell University Library, arXiv.org</publisher><subject>Classification ; Deep learning ; Entropy ; Machine learning ; Neural networks ; Regression ; Supervised learning ; Transformers</subject><ispartof>arXiv.org, 2024-03</ispartof><rights>2024. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>776,780</link.rule.ids></links><search><creatorcontrib>Farebrother, Jesse</creatorcontrib><creatorcontrib>Orbay, Jordi</creatorcontrib><creatorcontrib>Vuong, Quan</creatorcontrib><creatorcontrib>Adrien Ali Taïga</creatorcontrib><creatorcontrib>Chebotar, Yevgen</creatorcontrib><creatorcontrib>Xiao, Ted</creatorcontrib><creatorcontrib>Irpan, Alex</creatorcontrib><creatorcontrib>Levine, Sergey</creatorcontrib><creatorcontrib>Castro, Pablo Samuel</creatorcontrib><creatorcontrib>Faust, Aleksandra</creatorcontrib><creatorcontrib>Kumar, Aviral</creatorcontrib><creatorcontrib>Agarwal, Rishabh</creatorcontrib><title>Stop Regressing: Training Value Functions via Classification for Scalable Deep RL</title><title>arXiv.org</title><description>Value functions are a central component of deep reinforcement learning (RL). These functions, parameterized by neural networks, are trained using a mean squared error regression objective to match bootstrapped target values. However, scaling value-based RL methods that use regression to large networks, such as high-capacity Transformers, has proven challenging. This difficulty is in stark contrast to supervised learning: by leveraging a cross-entropy classification loss, supervised methods have scaled reliably to massive networks. Observing this discrepancy, in this paper, we investigate whether the scalability of deep RL can also be improved simply by using classification in place of regression for training value functions. We demonstrate that value functions trained with categorical cross-entropy significantly improves performance and scalability in a variety of domains. These include: single-task RL on Atari 2600 games with SoftMoEs, multi-task RL on Atari with large-scale ResNets, robotic manipulation with Q-transformers, playing Chess without search, and a language-agent Wordle task with high-capacity Transformers, achieving state-of-the-art results on these domains. Through careful analysis, we show that the benefits of categorical cross-entropy primarily stem from its ability to mitigate issues inherent to value-based RL, such as noisy targets and non-stationarity. Overall, we argue that a simple shift to training value functions with categorical cross-entropy can yield substantial improvements in the scalability of deep RL at little-to-no cost.</description><subject>Classification</subject><subject>Deep learning</subject><subject>Entropy</subject><subject>Machine learning</subject><subject>Neural networks</subject><subject>Regression</subject><subject>Supervised learning</subject><subject>Transformers</subject><issn>2331-8422</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>BENPR</sourceid><recordid>eNqNjMsKwjAURIMgWLT_cMF1oU3S-thWiws32uK2XEtaUkJSk8bvN4If4GqGOYdZkIgyliV7TumKxM6NaZrSYkfznEXkVs9mgrsYrHBO6uEIjUWpQ4MHKi-g8rqbpdEO3hKhVBi0Xnb43aA3FuoOFT6VgJMQ4em6IcselRPxL9dkW52b8pJM1ry8cHM7Gm91QC098IxzlrGC_Wd9AGxMPrE</recordid><startdate>20240306</startdate><enddate>20240306</enddate><creator>Farebrother, Jesse</creator><creator>Orbay, Jordi</creator><creator>Vuong, Quan</creator><creator>Adrien Ali Taïga</creator><creator>Chebotar, Yevgen</creator><creator>Xiao, Ted</creator><creator>Irpan, Alex</creator><creator>Levine, Sergey</creator><creator>Castro, Pablo Samuel</creator><creator>Faust, Aleksandra</creator><creator>Kumar, Aviral</creator><creator>Agarwal, Rishabh</creator><general>Cornell University Library, arXiv.org</general><scope>8FE</scope><scope>8FG</scope><scope>ABJCF</scope><scope>ABUWG</scope><scope>AFKRA</scope><scope>AZQEC</scope><scope>BENPR</scope><scope>BGLVJ</scope><scope>CCPQU</scope><scope>DWQXO</scope><scope>HCIFZ</scope><scope>L6V</scope><scope>M7S</scope><scope>PIMPY</scope><scope>PQEST</scope><scope>PQQKQ</scope><scope>PQUKI</scope><scope>PRINS</scope><scope>PTHSS</scope></search><sort><creationdate>20240306</creationdate><title>Stop Regressing: Training Value Functions via Classification for Scalable Deep RL</title><author>Farebrother, Jesse ; Orbay, Jordi ; Vuong, Quan ; Adrien Ali Taïga ; Chebotar, Yevgen ; Xiao, Ted ; Irpan, Alex ; Levine, Sergey ; Castro, Pablo Samuel ; Faust, Aleksandra ; Kumar, Aviral ; Agarwal, Rishabh</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-proquest_journals_29414431363</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Classification</topic><topic>Deep learning</topic><topic>Entropy</topic><topic>Machine learning</topic><topic>Neural networks</topic><topic>Regression</topic><topic>Supervised learning</topic><topic>Transformers</topic><toplevel>online_resources</toplevel><creatorcontrib>Farebrother, Jesse</creatorcontrib><creatorcontrib>Orbay, Jordi</creatorcontrib><creatorcontrib>Vuong, Quan</creatorcontrib><creatorcontrib>Adrien Ali Taïga</creatorcontrib><creatorcontrib>Chebotar, Yevgen</creatorcontrib><creatorcontrib>Xiao, Ted</creatorcontrib><creatorcontrib>Irpan, Alex</creatorcontrib><creatorcontrib>Levine, Sergey</creatorcontrib><creatorcontrib>Castro, Pablo Samuel</creatorcontrib><creatorcontrib>Faust, Aleksandra</creatorcontrib><creatorcontrib>Kumar, Aviral</creatorcontrib><creatorcontrib>Agarwal, Rishabh</creatorcontrib><collection>ProQuest SciTech Collection</collection><collection>ProQuest Technology Collection</collection><collection>Materials Science &amp; Engineering Collection</collection><collection>ProQuest Central (Alumni Edition)</collection><collection>ProQuest Central UK/Ireland</collection><collection>ProQuest Central Essentials</collection><collection>ProQuest Central</collection><collection>Technology Collection</collection><collection>ProQuest One Community College</collection><collection>ProQuest Central Korea</collection><collection>SciTech Premium Collection</collection><collection>ProQuest Engineering Collection</collection><collection>Engineering Database</collection><collection>Publicly Available Content Database</collection><collection>ProQuest One Academic Eastern Edition (DO NOT USE)</collection><collection>ProQuest One Academic</collection><collection>ProQuest One Academic UKI Edition</collection><collection>ProQuest Central China</collection><collection>Engineering Collection</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Farebrother, Jesse</au><au>Orbay, Jordi</au><au>Vuong, Quan</au><au>Adrien Ali Taïga</au><au>Chebotar, Yevgen</au><au>Xiao, Ted</au><au>Irpan, Alex</au><au>Levine, Sergey</au><au>Castro, Pablo Samuel</au><au>Faust, Aleksandra</au><au>Kumar, Aviral</au><au>Agarwal, Rishabh</au><format>book</format><genre>document</genre><ristype>GEN</ristype><atitle>Stop Regressing: Training Value Functions via Classification for Scalable Deep RL</atitle><jtitle>arXiv.org</jtitle><date>2024-03-06</date><risdate>2024</risdate><eissn>2331-8422</eissn><abstract>Value functions are a central component of deep reinforcement learning (RL). These functions, parameterized by neural networks, are trained using a mean squared error regression objective to match bootstrapped target values. However, scaling value-based RL methods that use regression to large networks, such as high-capacity Transformers, has proven challenging. This difficulty is in stark contrast to supervised learning: by leveraging a cross-entropy classification loss, supervised methods have scaled reliably to massive networks. Observing this discrepancy, in this paper, we investigate whether the scalability of deep RL can also be improved simply by using classification in place of regression for training value functions. We demonstrate that value functions trained with categorical cross-entropy significantly improves performance and scalability in a variety of domains. These include: single-task RL on Atari 2600 games with SoftMoEs, multi-task RL on Atari with large-scale ResNets, robotic manipulation with Q-transformers, playing Chess without search, and a language-agent Wordle task with high-capacity Transformers, achieving state-of-the-art results on these domains. Through careful analysis, we show that the benefits of categorical cross-entropy primarily stem from its ability to mitigate issues inherent to value-based RL, such as noisy targets and non-stationarity. Overall, we argue that a simple shift to training value functions with categorical cross-entropy can yield substantial improvements in the scalability of deep RL at little-to-no cost.</abstract><cop>Ithaca</cop><pub>Cornell University Library, arXiv.org</pub><oa>free_for_read</oa></addata></record>
fulltext fulltext
identifier EISSN: 2331-8422
ispartof arXiv.org, 2024-03
issn 2331-8422
language eng
recordid cdi_proquest_journals_2941443136
source Free E- Journals
subjects Classification
Deep learning
Entropy
Machine learning
Neural networks
Regression
Supervised learning
Transformers
title Stop Regressing: Training Value Functions via Classification for Scalable Deep RL
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-28T20%3A19%3A10IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-proquest&rft_val_fmt=info:ofi/fmt:kev:mtx:book&rft.genre=document&rft.atitle=Stop%20Regressing:%20Training%20Value%20Functions%20via%20Classification%20for%20Scalable%20Deep%20RL&rft.jtitle=arXiv.org&rft.au=Farebrother,%20Jesse&rft.date=2024-03-06&rft.eissn=2331-8422&rft_id=info:doi/&rft_dat=%3Cproquest%3E2941443136%3C/proquest%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=2941443136&rft_id=info:pmid/&rfr_iscdi=true