東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁

切換: 標籤 | MARC模式 | ISBD

Towards Alignment of Reinforcement L...

Satija, Harsh,

FindBook

Google Book

Amazon

博客來

Towards Alignment of Reinforcement Learning Agents; for Consideration of Safety, Robustness and Fairness /

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Towards Alignment of Reinforcement Learning Agents; for Consideration of Safety, Robustness and Fairness // Harsh Satija.
作者:	Satija, Harsh,
面頁冊數:	1 electronic resource (239 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
Contained By:	Dissertations Abstracts International86-03B.
標題:	Pessimism. -
電子資源:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31339097
ISBN:	9798384219415

Towards Alignment of Reinforcement Learning Agents; for Consideration of Safety, Robustness and Fairness /
Satija, Harsh,

Towards Alignment of Reinforcement Learning Agents; for Consideration of Safety, Robustness and Fairness /Harsh Satija. - 1 electronic resource (239 pages)

Source: Dissertations Abstracts International, Volume: 86-03, Section: B.

Reinforcement Learning (RL) has emerged as the standard paradigm for sequential decision-making and a framework for general intelligence. At its core, the RL problem is one of trial-and-error, where the agent explores all sorts of behaviours, even harmful ones, in pursuit of the one that exploits the task objective. In this dissertation, I investigate how we can better align the RL agents so that they do not cause undesirable behaviour or unintended consequences. I first study the settings where the practitioner explicitly knows how to define unwanted behaviour based on the notions of safety or fairness. For such scenarios, I present algorithms that allow the RL agents to avoid undesirable behaviours while learning to solve the given task. Then, I investigate the settings where we want the RL agents to be resilient, either from the practitioners regarding misspecified task objectives or due to slight variations in the learning environments. Finally, I present algorithms that allow practitioners to experiment with reward design strategies in high-stake settings without worrying about the unintended consequences of the resulting solution strategy.

English

ISBN: 9798384219415Subjects--Topical Terms:

565078
Pessimism.

Towards Alignment of Reinforcement Learning Agents; for Consideration of Safety, Robustness and Fairness /
LDR:03819nmm a22003853i 4500 001 2438233
005 20251210113950.5
006 m o d
007 cr|nu||||||||
008 261011s2024 miu||||||m |||||||eng d
020 $a 9798384219415
035 $a (MiAaPQD)AAI31339097
035 $a (MiAaPQD)McGill5x21tn08n
035 $a AAI31339097
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Satija, Harsh, $e author. $3 3829476
245 1 0 $a Towards Alignment of Reinforcement Learning Agents; for Consideration of Safety, Robustness and Fairness / $c Harsh Satija.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 1 electronic resource (239 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
500 $a Advisors: Pineau, Joelle.
502 $b Ph.D. $c McGill University (Canada) $d 2024.
520 $a Reinforcement Learning (RL) has emerged as the standard paradigm for sequential decision-making and a framework for general intelligence. At its core, the RL problem is one of trial-and-error, where the agent explores all sorts of behaviours, even harmful ones, in pursuit of the one that exploits the task objective. In this dissertation, I investigate how we can better align the RL agents so that they do not cause undesirable behaviour or unintended consequences. I first study the settings where the practitioner explicitly knows how to define unwanted behaviour based on the notions of safety or fairness. For such scenarios, I present algorithms that allow the RL agents to avoid undesirable behaviours while learning to solve the given task. Then, I investigate the settings where we want the RL agents to be resilient, either from the practitioners regarding misspecified task objectives or due to slight variations in the learning environments. Finally, I present algorithms that allow practitioners to experiment with reward design strategies in high-stake settings without worrying about the unintended consequences of the resulting solution strategy.
520 $a L'apprentissage par renforcement (RL) s'est impose comme le paradigme standard de la prise de decision sequentielle et un cadre pour l'intelligence generale. Fondamentalement, le probleme de l'apprentissage par renforcement est un probleme d'essais et d'erreurs, ou l'agent explore toutes sortes de comportements, meme nuisibles, a la recherche de celui qui permet de reussir au mieux l'objectif de la tache. Dans cette these, j'etudie comment nous pouvons mieux aligner les agents RL afin qu'ils ne provoquent pas de comportements indesirables ou de consequences involontaires. J'etudie d'abord les contextes ou le praticien sait explicitement comment definir un comportement indesirable sur la base de notions de securite ou d'equite. Pour ces scenarios, je presente des algorithmes qui permettent aux agents RL d'eviter ces comportements indesirables tout en apprenant a resoudre la tache donnee. Ensuite, j'etudie les contextes dans lesquels nous voulons que les agents RL soient resilients, que ce soit face aux praticiens concernant des objectifs de taches mal specifies ou en raison de legeres variations dans les environnements d'apprentissage. Enfin, je presente des algorithmes qui permettent aux praticiens d'experimenter des strategies de conception de recompenses dans des contextes a fort enjeu sans s'inquieter des consequences involontaires de solution resultante.
546 $a English
590 $a School code: 0781
650 4 $a Pessimism. $3 565078
650 4 $a Linear programming. $3 560448
650 4 $a Dynamic programming. $3 641303
650 4 $a Decision making. $3 517204
650 4 $a Markov analysis. $3 3562906
650 4 $a Computer science. $3 523869
690 $a 0984
690 $a 0796
710 2 $a McGill University (Canada). $e degree granting institution. $3 3759190
720 1 $a Pineau, Joelle $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-03B.
790 $a 0781
791 $a Ph.D.
792 $a 2024
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31339097