<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Publishing DTD v1.3 20210610//EN" "JATS-journalpublishing1-3.dtd">
<article article-type="research-article" dtd-version="1.3" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xml:lang="ru"><front><journal-meta><journal-id journal-id-type="publisher-id">novtexmech</journal-id><journal-title-group><journal-title xml:lang="ru">Мехатроника, автоматизация, управление</journal-title><trans-title-group xml:lang="en"><trans-title>Mekhatronika, Avtomatizatsiya, Upravlenie</trans-title></trans-title-group></journal-title-group><issn pub-type="ppub">1684-6427</issn><issn pub-type="epub">2619-1253</issn><publisher><publisher-name>Commercial Publisher «New Technologies»</publisher-name></publisher></journal-meta><article-meta><article-id pub-id-type="doi">10.17587/mau.23.414-419</article-id><article-id custom-type="elpub" pub-id-type="custom">novtexmech-1231</article-id><article-categories><subj-group subj-group-type="heading"><subject>Research Article</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="ru"><subject>СИСТЕМНЫЙ АНАЛИЗ, УПРАВЛЕНИЕ И ОБРАБОТКА ИНФОРМАЦИИ</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="en"><subject>SYSTEM ANALYSIS, CONTROL AND INFORMATION PROCESSING</subject></subj-group></article-categories><title-group><article-title>Метод преодоления катастрофического забывания при мультизадачном обучении с подкреплением</article-title><trans-title-group xml:lang="en"><trans-title>A Method for Catastrophic Forgetting Prevention during Multitasking Reinforcement Learning</trans-title></trans-title-group></title-group><contrib-group><contrib contrib-type="author" corresp="yes"><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Аглюков</surname><given-names>И. Н.</given-names></name><name name-style="western" xml:lang="en"><surname>Agliukov</surname><given-names>I. N.</given-names></name></name-alternatives><bio xml:lang="ru"><p>студент, </p><p>Москва</p></bio><bio xml:lang="en"><p>Moscow, 101000</p></bio><email xlink:type="simple">ildariwe@gmail.com</email><xref ref-type="aff" rid="aff-1"/></contrib><contrib contrib-type="author" corresp="yes"><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Святов</surname><given-names>К. В.</given-names></name><name name-style="western" xml:lang="en"><surname>Sviatov</surname><given-names>K. V.</given-names></name></name-alternatives><bio xml:lang="ru"><p>канд. техн. наук,</p><p>г. Ульяновск</p></bio><bio xml:lang="en"><p>Ulyanovsk, 432027</p></bio><email xlink:type="simple">k.svyatov@ulstu.ru</email><xref ref-type="aff" rid="aff-2"/></contrib><contrib contrib-type="author" corresp="yes"><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Сухов</surname><given-names>С. В.</given-names></name><name name-style="western" xml:lang="en"><surname>Sukhov</surname><given-names>S. V.</given-names></name></name-alternatives><bio xml:lang="ru"><p>канд. физ.-мат. наук, </p><p>г. Ульяновск</p></bio><bio xml:lang="en"><p>Сand. Sc.,</p><p>Ulyanovsk, 432071</p></bio><email xlink:type="simple">ssukhov@ulireran.ru</email><xref ref-type="aff" rid="aff-3"/></contrib></contrib-group><aff-alternatives id="aff-1"><aff xml:lang="ru"><institution>Высшая школа экономики</institution><country>Россия</country></aff><aff xml:lang="en"><institution>National Research University Higher School of Economics</institution><country>Russian Federation</country></aff></aff-alternatives><aff-alternatives id="aff-2"><aff xml:lang="ru"><institution>Ульяновский государственный технический университет</institution><country>Россия</country></aff><aff xml:lang="en"><institution>Ulyanovsk State Technical University</institution><country>Russian Federation</country></aff></aff-alternatives><aff-alternatives id="aff-3"><aff xml:lang="ru"><institution>УФИРЭ им. В. А. Котельникова РАН</institution><country>Россия</country></aff><aff xml:lang="en"><institution>Ulyanovsk Branch of Kotelnikov Institute of Radioengineering and Electronics of Russian Academy of Sciences</institution><country>Russian Federation</country></aff></aff-alternatives><pub-date pub-type="collection"><year>2022</year></pub-date><pub-date pub-type="epub"><day>13</day><month>08</month><year>2022</year></pub-date><volume>23</volume><issue>8</issue><fpage>414</fpage><lpage>419</lpage><permissions><copyright-statement>Copyright &amp;#x00A9; Commercial Publisher «New Technologies», 2022</copyright-statement><copyright-year>2022</copyright-year><copyright-holder xml:lang="ru">Commercial Publisher «New Technologies»</copyright-holder><copyright-holder xml:lang="en">Commercial Publisher «New Technologies»</copyright-holder><license xlink:href="https://mech.novtex.ru/jour/about/submissions#copyrightNotice" xlink:type="simple"><license-p>https://mech.novtex.ru/jour/about/submissions#copyrightNotice</license-p></license></permissions><self-uri xlink:href="https://mech.novtex.ru/jour/article/view/1231">https://mech.novtex.ru/jour/article/view/1231</self-uri><abstract><p>Принцип обучения с подкреплением основан на взаимодействии агента с окружением в целях максимизации своей награды. Обучение с подкреплением показывает очень хорошие результаты в решении различных задач управления. Тем не менее, попытки обучить интеллектуального агента эффективно решать несколько задач страдают от проблемы так называемого "катастрофического забывания". Полученные агентом знания об одной задаче вытесняются информацией в ходе выработки правильной стратегии для другой. Одним из методов предотвращения катастрофического забывания при многозадачном обучении является обучение агента на сохраненных в буфере опыта ранее встреченных состояниях. Разработанный нами метод позволяет обучить агента тому, как эффективно вести себя в нескольких средах одновременно на основе обмена опытом с агентами-учителями, используя буфер опыта. Обмен опытом основан на распространенном в глубоком обучении подходе, называемом дистилляцией знаний. Дистилляция знаний позволила свести задачу с подкреплением к задаче обучения с учителем. В ходе исследований были протестированы и выбраны максимально успешные сочетания различных функций потерь и способов преобразования выходных слоев нейросетей. Метод дистилляции знаний требует хранения огромного буфера состояний. Предложены несколько методик оптимизации хранения буфера: использование части буфера и сжатие состояний во внутреннее представление нейросети с помощью автокодировщика. В качестве тестового окружения для экспериментов использовались игры Atari. </p></abstract><trans-abstract xml:lang="en"><p>Reinforcement learning is based on a principle of an agent interacting with an environment in order to maximize the amount of reward. Reinforcement learning shows amazing results in solving various control problems. However, the attempts to train a multitasking agent suffer from the problem of so-called "catastrophic forgetting": the knowledge gained by the agent about one task is erased during developing the correct strategy to solve another task. One of the methods to fight catastrophic forgetting during multitask learning assumes storing previously encountered states in, the so-called, experience replay buffer. We developed the method allowing a student agent to exchange an experience with teacher agents using an experience replay buffer. The procedure of experience exchange allowed the student to behave effectively in several environments simultaneously. The experience exchange was based on knowledge distillation that allowed to reduce the off-policy reinforcement learning problem to the supervised learning task. We tested several combinations of loss functions and output transforming functions. Distillation of knowledge requires a massive experience replay buffer. Several solutions to the problems of optimizing the size of the experience replay buffer are suggested. The first approach is based on the use of a subset of the whole buffer; the second approach uses the autoencoder as a tool to convert states to the latent space. Although our methods can be applied to a wide range of problems, we use Atari games as a testing environment to demonstrate the methods. </p></trans-abstract><kwd-group xml:lang="ru"><kwd>обучение с подкреплением</kwd><kwd>автономное обучение с подкреплением</kwd><kwd>мультизадачное обучение</kwd><kwd>обмен опытом</kwd><kwd>буфер опыта</kwd><kwd>дистилляция стратегии</kwd><kwd>клонирование поведения</kwd><kwd>имитационное обучение</kwd><kwd>катастрофическое забывание</kwd><kwd>непрерывное обучение</kwd></kwd-group><kwd-group xml:lang="en"><kwd>reinforcement learning</kwd><kwd>offline reinforcement learning</kwd><kwd>multitask learning</kwd><kwd>experience exchange</kwd><kwd>experience replay buffer</kwd><kwd>policy distillation</kwd><kwd>behavior cloning</kwd><kwd>imitation learning</kwd><kwd>catastrophic forgetting</kwd><kwd>continuous learning</kwd></kwd-group><funding-group><funding-statement xml:lang="ru">Исследование выполнено при финансовой поддержке РФФИ и Правительства Ульяновской области (проект № 18-47-732006 ).</funding-statement><funding-statement xml:lang="en">The study was financially supported by the Russian Foundation for Basic Research and the Government of the Ulyanovsk Region (Project No. 18-47-732006).</funding-statement></funding-group></article-meta></front><back><ref-list><title>References</title><ref id="cit1"><label>1</label><citation-alternatives><mixed-citation xml:lang="ru">Shmygun A. A, Ermolaeva L. V., Zakharov N. V. Obuchenie s podkrepleniem, Novaya Nauka: sovremennoe sostoyanie i puti razvitiya, 2016, no. 12-3, pp. 189—191, available at: https://elibrary.ru/download/elibrary_27724493_81982095.pdf (in Russian).</mixed-citation><mixed-citation xml:lang="en">Shmygun A. A, Ermolaeva L. V., Zakharov N. V. Obuchenie s podkrepleniem, Novaya Nauka: sovremennoe sostoyanie i puti razvitiya, 2016, no. 12-3, pp. 189—191, available at: https://elibrary.ru/download/elibrary_27724493_81982095.pdf (in Russian).</mixed-citation></citation-alternatives></ref><ref id="cit2"><label>2</label><citation-alternatives><mixed-citation xml:lang="ru">Ecoffet A., Huizinga J., Lehman J. J., Stanley K. O., Clune J. First return, then explore, Nature, 2021, vol. 590, no. 7847, pp. 580—586, DOI: 10.1038/s41586-020-03157-9.</mixed-citation><mixed-citation xml:lang="en">Ecoffet A., Huizinga J., Lehman J. J., Stanley K. O., Clune J. First return, then explore, Nature, 2021, vol. 590, no. 7847, pp. 580—586, DOI: 10.1038/s41586-020-03157-9.</mixed-citation></citation-alternatives></ref><ref id="cit3"><label>3</label><citation-alternatives><mixed-citation xml:lang="ru">Kalashnikov D., Irpan A., Pastor P., Ibarz J., Herzog A., Jang E., Quillen D., Holly E., Kalakrishnan M., Vanhoucke V., Levine S. Qt-opt: Scalable deep reinforcement learning for vision based robotic manipulation, arXiv preprint arXiv:1806.10293, 2018, available at: https://arxiv.org/abs/1806.10293.</mixed-citation><mixed-citation xml:lang="en">Kalashnikov D., Irpan A., Pastor P., Ibarz J., Herzog A., Jang E., Quillen D., Holly E., Kalakrishnan M., Vanhoucke V., Levine S. Qt-opt: Scalable deep reinforcement learning for vision based robotic manipulation, arXiv preprint arXiv:1806.10293, 2018, available at: https://arxiv.org/abs/1806.10293.</mixed-citation></citation-alternatives></ref><ref id="cit4"><label>4</label><citation-alternatives><mixed-citation xml:lang="ru">Da Silva F. L., Taylor M. E., Costa A. H. R. Autonomously reusing knowledge in multiagent reinforcement learning, Proc. 27th Int. Joint Conf. on Artificial Intelligence, 2018, pp. 5487—5493, available at: https://www.ijcai.org/proceedings/2018/0774.pdf.</mixed-citation><mixed-citation xml:lang="en">Da Silva F. L., Taylor M. E., Costa A. H. R. Autonomously reusing knowledge in multiagent reinforcement learning, Proc. 27th Int. Joint Conf. on Artificial Intelligence, 2018, pp. 5487—5493, available at: https://www.ijcai.org/proceedings/2018/0774.pdf.</mixed-citation></citation-alternatives></ref><ref id="cit5"><label>5</label><citation-alternatives><mixed-citation xml:lang="ru">Koroteev M. V. Obzor nekotoryh sovremennyh tendentsyj v tehnologiyah mashinnogo obucheniya, E-Management, 2018, pp. 30—31 (in Russian).</mixed-citation><mixed-citation xml:lang="en">Koroteev M. V. Obzor nekotoryh sovremennyh tendentsyj v tehnologiyah mashinnogo obucheniya, E-Management, 2018, pp. 30—31 (in Russian).</mixed-citation></citation-alternatives></ref><ref id="cit6"><label>6</label><citation-alternatives><mixed-citation xml:lang="ru">Lesort T., Lomonaco V., Stoian A., Maltoni D., Filliat D., Díaz-Rodríguez N. Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges, Information Fusion, 2020, vol. 58, pp. 52—68, DOI: 10.1016/j.inffus.2019.12.004.</mixed-citation><mixed-citation xml:lang="en">Lesort T., Lomonaco V., Stoian A., Maltoni D., Filliat D., Díaz-Rodríguez N. Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges, Information Fusion, 2020, vol. 58, pp. 52—68, DOI: 10.1016/j.inffus.2019.12.004.</mixed-citation></citation-alternatives></ref><ref id="cit7"><label>7</label><citation-alternatives><mixed-citation xml:lang="ru">Ross S., Gordon G., Bagnell A. A reduction of imitation learning and structured prediction to no-regret online learning, Journal of Machine Learning Research, 2011, vol. 15, pp. 627—635, available at: http://proceedings.mlr.press/v15/ross11a.</mixed-citation><mixed-citation xml:lang="en">Ross S., Gordon G., Bagnell A. A reduction of imitation learning and structured prediction to no-regret online learning, Journal of Machine Learning Research, 2011, vol. 15, pp. 627—635, available at: http://proceedings.mlr.press/v15/ross11a.</mixed-citation></citation-alternatives></ref><ref id="cit8"><label>8</label><citation-alternatives><mixed-citation xml:lang="ru">Parisotto E., Lei Ba J., Salakhutdinov R. Actor-mimic: Deep multitask and transfer reinforcement learning, arXiv preprint arXiv:1511.06342. 2015, available at: https://arxiv.org/abs/1511.06342.</mixed-citation><mixed-citation xml:lang="en">Parisotto E., Lei Ba J., Salakhutdinov R. Actor-mimic: Deep multitask and transfer reinforcement learning, arXiv preprint arXiv:1511.06342. 2015, available at: https://arxiv.org/abs/1511.06342.</mixed-citation></citation-alternatives></ref><ref id="cit9"><label>9</label><citation-alternatives><mixed-citation xml:lang="ru">Teh Y. W., Bapst V., Czarnecki W. M., Quan J., Kirkpatrick J., Hadsell R., Heess N., Pascanu R. Distral: Robust multitask reinforcement learning, Proc. 31st Int. Conf. on Neural Information Processing Systems, 2017, pp. 4499—4509, available at: https://proceedings.neurips.cc/paper/2017/hash/0abdc563a06105a ee3c6136871c9f4d1-Abstract.html.</mixed-citation><mixed-citation xml:lang="en">Teh Y. W., Bapst V., Czarnecki W. M., Quan J., Kirkpatrick J., Hadsell R., Heess N., Pascanu R. Distral: Robust multitask reinforcement learning, Proc. 31st Int. Conf. on Neural Information Processing Systems, 2017, pp. 4499—4509, available at: https://proceedings.neurips.cc/paper/2017/hash/0abdc563a06105a ee3c6136871c9f4d1-Abstract.html.</mixed-citation></citation-alternatives></ref><ref id="cit10"><label>10</label><citation-alternatives><mixed-citation xml:lang="ru">Rusu A. A., Colmenarejo S. G., Gulcehre C., Desjardins G., Kirkpatrick J., Pascanu R., Mnih V., Kavukcuoglu K., Hadsell R. Policy distillation, arXiv preprint arXiv:1511.06295. 2015. URL: https://arxiv.org/abs/1511.06295.</mixed-citation><mixed-citation xml:lang="en">Rusu A. A., Colmenarejo S. G., Gulcehre C., Desjardins G., Kirkpatrick J., Pascanu R., Mnih V., Kavukcuoglu K., Hadsell R. Policy distillation, arXiv preprint arXiv:1511.06295. 2015. URL: https://arxiv.org/abs/1511.06295.</mixed-citation></citation-alternatives></ref><ref id="cit11"><label>11</label><citation-alternatives><mixed-citation xml:lang="ru">Levine S., Kumar A., Tucker G., Fu J. Offline reinforcement learning: Tutorial, review, and perspectives on open problems, arXiv preprint arXiv:2005.01643. 2020. URL: https://arxiv.org/abs/2005.01643.</mixed-citation><mixed-citation xml:lang="en">Levine S., Kumar A., Tucker G., Fu J. Offline reinforcement learning: Tutorial, review, and perspectives on open problems, arXiv preprint arXiv:2005.01643. 2020. URL: https://arxiv.org/abs/2005.01643.</mixed-citation></citation-alternatives></ref><ref id="cit12"><label>12</label><citation-alternatives><mixed-citation xml:lang="ru">Ballard D. Modular learning in neural networks, Proc. 6th National Conf. on Artificial Intelligence, 1987, vol. 1, pp. 279—284, available at: https://www.aaai.org/Library/AAAI/1987/aaai87-050.php.</mixed-citation><mixed-citation xml:lang="en">Ballard D. Modular learning in neural networks, Proc. 6th National Conf. on Artificial Intelligence, 1987, vol. 1, pp. 279—284, available at: https://www.aaai.org/Library/AAAI/1987/aaai87-050.php.</mixed-citation></citation-alternatives></ref><ref id="cit13"><label>13</label><citation-alternatives><mixed-citation xml:lang="ru">Akinina N. V., Akinin M. V., Sokolova A. V., Nikiforov M. B., Taganov A. I. Avoenkoder: podhod k ponizheniyu razmernosti vektornogo prostranstva s kontroliruemoj poterej informatsii // Izvestiya Tul’skogo gosudarstvennogo universiteta. Tehnicheskie nauki. 2016. no. 9. pp. 3—12, available at: https://elibrary.ru/download/elibrary_27277969_89179513.pdf (in Russian).</mixed-citation><mixed-citation xml:lang="en">Akinina N. V., Akinin M. V., Sokolova A. V., Nikiforov M. B., Taganov A. I. Avoenkoder: podhod k ponizheniyu razmernosti vektornogo prostranstva s kontroliruemoj poterej informatsii // Izvestiya Tul’skogo gosudarstvennogo universiteta. Tehnicheskie nauki. 2016. no. 9. pp. 3—12, available at: https://elibrary.ru/download/elibrary_27277969_89179513.pdf (in Russian).</mixed-citation></citation-alternatives></ref><ref id="cit14"><label>14</label><citation-alternatives><mixed-citation xml:lang="ru">McCloskey M., Cohen N. J. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem, Psychology of learning and motivation, 1989, vol. 24, pp. 109—165, DOI: 10.1016/S0079-7421(08)60536-8.</mixed-citation><mixed-citation xml:lang="en">McCloskey M., Cohen N. J. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem, Psychology of learning and motivation, 1989, vol. 24, pp. 109—165, DOI: 10.1016/S0079-7421(08)60536-8.</mixed-citation></citation-alternatives></ref><ref id="cit15"><label>15</label><citation-alternatives><mixed-citation xml:lang="ru">Sukhov S., Leontev M., Miheev A., Sviatov K. Prevention of catastrophic interference and imposing active forgetting with generative methods, Neurocomputing, 2020, vol. 400, pp.73—85, DOI: 10.1016/j.neucom.2020.03.024.</mixed-citation><mixed-citation xml:lang="en">Sukhov S., Leontev M., Miheev A., Sviatov K. Prevention of catastrophic interference and imposing active forgetting with generative methods, Neurocomputing, 2020, vol. 400, pp.73—85, DOI: 10.1016/j.neucom.2020.03.024.</mixed-citation></citation-alternatives></ref><ref id="cit16"><label>16</label><citation-alternatives><mixed-citation xml:lang="ru">Mnih V., Kavukcuoglu K., Silver D., Rusu A. A., Veness J., Bellemare M. G., Graves A., Riedmiller M., Fidjeland A. K., Ostrovski G., Petersen S., Beattie C., Sadik A., Antonoglou I., King H., Kumaran D., Wierstra D., Legg S., Hassabis D. Human-level control through deep reinforcement lear ning, Nature, 2015, vol. 518, no. 7540, pp. 529—533, DOI: 10.1038/nature14236.</mixed-citation><mixed-citation xml:lang="en">Mnih V., Kavukcuoglu K., Silver D., Rusu A. A., Veness J., Bellemare M. G., Graves A., Riedmiller M., Fidjeland A. K., Ostrovski G., Petersen S., Beattie C., Sadik A., Antonoglou I., King H., Kumaran D., Wierstra D., Legg S., Hassabis D. Human-level control through deep reinforcement lear ning, Nature, 2015, vol. 518, no. 7540, pp. 529—533, DOI: 10.1038/nature14236.</mixed-citation></citation-alternatives></ref><ref id="cit17"><label>17</label><citation-alternatives><mixed-citation xml:lang="ru">Wang Z., de Freitas N., Lanctot M. Dueling Network Architectures for Deep Reinforcement Learning, Int. Conf. on Machine Learning, 2015, pp. 1995—2003, available at: http://proceedings.mlr.press/v48/wangf16.html.</mixed-citation><mixed-citation xml:lang="en">Wang Z., de Freitas N., Lanctot M. Dueling Network Architectures for Deep Reinforcement Learning, Int. Conf. on Machine Learning, 2015, pp. 1995—2003, available at: http://proceedings.mlr.press/v48/wangf16.html.</mixed-citation></citation-alternatives></ref><ref id="cit18"><label>18</label><citation-alternatives><mixed-citation xml:lang="ru">Biryukova V. A. Tehnologiya distillyatsii znanij dlya obucheniya nejronnyh setej na primere zadachi binarnoj klassifikacii, Intellektual’nye sistemy. Teoriya i prilozheniya, 2020, vol. 24, no. 2, pp. 23—52, available at: http://intsysjournal.ru/pdfs/24-2/Birukova.pdf (in Russian).</mixed-citation><mixed-citation xml:lang="en">Biryukova V. A. Tehnologiya distillyatsii znanij dlya obucheniya nejronnyh setej na primere zadachi binarnoj klassifikacii, Intellektual’nye sistemy. Teoriya i prilozheniya, 2020, vol. 24, no. 2, pp. 23—52, available at: http://intsysjournal.ru/pdfs/24-2/Birukova.pdf (in Russian).</mixed-citation></citation-alternatives></ref><ref id="cit19"><label>19</label><citation-alternatives><mixed-citation xml:lang="ru">Brockman G., Cheung V., Pettersson L., Schneider J., Schulman J., Tang J., Zaremba W. Openai gym, arXiv preprint arXiv:1606.01540, 2016, available at: https://arxiv.org/abs/1606.01540.</mixed-citation><mixed-citation xml:lang="en">Brockman G., Cheung V., Pettersson L., Schneider J., Schulman J., Tang J., Zaremba W. Openai gym, arXiv preprint arXiv:1606.01540, 2016, available at: https://arxiv.org/abs/1606.01540.</mixed-citation></citation-alternatives></ref></ref-list><fn-group><fn fn-type="conflict"><p>The authors declare that there are no conflicts of interest present.</p></fn></fn-group></back></article>
