서브메뉴
검색
Efficiently Imitating Human Movement in Counter-Strike
Efficiently Imitating Human Movement in Counter-Strike
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211153056
- ISBN
- 9798346382850
- DDC
- 355
- 서명/저자
- Efficiently Imitating Human Movement in Counter-Strike
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 106 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-06, Section: B.
- 주기사항
- Advisor: Fatahalian, Kayvon.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약Human-like agents have the potential to drastically improve multiplayer, first-person shooter (FPS) games. They can serve as engaging teammates, useful practice partners, and anti-social behavior detectors. However, it is difficult to create a multiplayer FPS agent that replicates human behavior, and in particular human movement. Despite game developers' efforts for decades, agents either struggle with the wide range of possible game situations in a multiplayer FPS, or are too computationally inefficient to deploy in a commercial title.This dissertation contributes a machine learning (ML)-based multiplayer FPS agent that has some of the most human-like behavior demonstrated to date while also satisfying games' performance constraints. Previous game developers avoided ML agent techniques due to the computational requirements. In order to make human-like ML tractable for real-time use in a commercial game, we made three key design decisions. First, we use an imitation learning approach to train the agent. Imitation is the most direct approach for creating human-like agents, and multiplayer FPS titles generate large datasets of demonstrations to imitate. Second, long-term human-like behavior emerges from our ML model's short-term predictions. Predicting only the next few actions enables us to automatically generate large collections of labels and utilize a simple, supervised training process. Finally, we only utilize learning where necessary. Our agent's hybrid architecture utilizes rule-based behavior generators where possible, and the ML model's game state input is in a symbolic format that can be efficiently processed.We utilize these principles to create a complete agent system. We will describe the system's four components. First, a dataset curation pipeline for creating a large-scale dataset of human movement. Second, an efficient, transformer-based movement model trained to imitate the dataset. Third, a complete agent, known as MLMove, that utilizes the learned movement model to play the multiplayer FPS game Counter-Strike. Finally, we evaluate if MLMove is human-like. Since "humanness" is difficult to evaluate, we present a suite of evaluations, including a user study and large-scale analytics, demonstrating that MLMove's behavior is more human-like than strong baselines including industry-standard and expert-crafted agents. We discuss our agent's limitations and the impact of our work on the wider field of video game agents.
- 일반주제명
- Defense
- 일반주제명
- Computer engineering
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-06B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017164866
■00520250211153056
■006m o d
■007cr#unu||||||||
■020 ▼a9798346382850
■035 ▼a(MiAaPQ)AAI31643418
■035 ▼a(MiAaPQ)Stanfordyz173qh1790
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a355
■1001 ▼aDurst, David Benjamin.
■24510▼aEfficiently Imitating Human Movement in Counter-Strike
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a106 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-06, Section: B.
■500 ▼aAdvisor: Fatahalian, Kayvon.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aHuman-like agents have the potential to drastically improve multiplayer, first-person shooter (FPS) games. They can serve as engaging teammates, useful practice partners, and anti-social behavior detectors. However, it is difficult to create a multiplayer FPS agent that replicates human behavior, and in particular human movement. Despite game developers' efforts for decades, agents either struggle with the wide range of possible game situations in a multiplayer FPS, or are too computationally inefficient to deploy in a commercial title.This dissertation contributes a machine learning (ML)-based multiplayer FPS agent that has some of the most human-like behavior demonstrated to date while also satisfying games' performance constraints. Previous game developers avoided ML agent techniques due to the computational requirements. In order to make human-like ML tractable for real-time use in a commercial game, we made three key design decisions. First, we use an imitation learning approach to train the agent. Imitation is the most direct approach for creating human-like agents, and multiplayer FPS titles generate large datasets of demonstrations to imitate. Second, long-term human-like behavior emerges from our ML model's short-term predictions. Predicting only the next few actions enables us to automatically generate large collections of labels and utilize a simple, supervised training process. Finally, we only utilize learning where necessary. Our agent's hybrid architecture utilizes rule-based behavior generators where possible, and the ML model's game state input is in a symbolic format that can be efficiently processed.We utilize these principles to create a complete agent system. We will describe the system's four components. First, a dataset curation pipeline for creating a large-scale dataset of human movement. Second, an efficient, transformer-based movement model trained to imitate the dataset. Third, a complete agent, known as MLMove, that utilizes the learned movement model to play the multiplayer FPS game Counter-Strike. Finally, we evaluate if MLMove is human-like. Since "humanness" is difficult to evaluate, we present a suite of evaluations, including a user study and large-scale analytics, demonstrating that MLMove's behavior is more human-like than strong baselines including industry-standard and expert-crafted agents. We discuss our agent's limitations and the impact of our work on the wider field of video game agents.
■590 ▼aSchool code: 0212.
■650 4▼aDefense
■650 4▼aComputer engineering
■653 ▼aHuman-like agents
■653 ▼aFirst-person shooter games
■690 ▼a0464
■690 ▼a0800
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-06B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17164866▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


