서브메뉴
검색
An α-Potential Game Framework for Non-Cooperative Dynamic Games: Theory and Algorithms
An α-Potential Game Framework for Non-Cooperative Dynamic Games: Theory and Algorithms
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104706
- ISBN
- 9798288866678
- DDC
- 519
- 저자명
- Li, Xinyu.
- 서명/저자
- An α-Potential Game Framework for Non-Cooperative Dynamic Games: Theory and Algorithms
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 164 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Guo, Xin.
- 학위논문주기
- Thesis (D.Eng.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약Multi-agent systems naturally arise in many real-world scenarios where multiple decision-makers interact, such as autonomous driving, network design, and financial markets. Analyzing multi-agent non-cooperative games is inherently challenging due to the asymmetry among players and the diverse structures of the games. This thesis introduces a unified framework to analyze N-player non-cooperative games in dynamic settings, considering both discrete-time and continuous-time state transitions. Additionally, efficient reinforcement learning (RL) algorithms and stochastic control techniques are proposed to identify strategies for each player that lead to approximate Nash equilibria (NE).The first part of this thesis introduces and analyzes a general class of dynamic N-player non-cooperative games called α-potential games. In this framework, the change in a player's objective function resulting from a unilateral deviation from her strategy is equal to the change in a common function, called the α-potential function, up to an error α. The existence of an α-potential function simplifies the challenging task of finding α-Nash equilibria in dynamic games to minimizing the α-potential function, as the optimizer of the α-potential function is shown to be an α-Nash equilibrium of the game.In Chapter 2, we focus on Markov games with finite state space, finite action space, and Markovian policy. The state transition follows a discrete-time Markov decision process. In this case, we establish the existence of an associated α-potential function. Additionally, we provide a semi-infinite linear program to find α and its corresponding α-potential function for any Markov game. We study two important classes of practically significant Markov games, Markov congestion games and the perturbed Markov team games, via the framework of Markov α-potential games, with explicit characterization of an upper bound for α and its relation to game parameters. Furthermore, we study two equilibrium approximation algorithms, namely the projected gradient-ascent algorithm and the sequential maximum improvement algorithm, along with their Nash regret analysis, and corroborate the results with numerical experiments using model-free RL algorithms.In Chapter 3, we study dynamic games with continuous-time state transitions, focusing on general classes of states, actions, and controls/policies, with a particular emphasis on stochastic differential games. An analytical characterization of the α-potential function is established, with α represented in terms of the magnitude of the asymmetry of the second-order derivatives of the players' objective functions. For stochastic differential games in which the state dynamic is a controlled diffusion, α is characterized in terms of the number of players, the choice of admissible strategies, and the intensity of interactions and the level of heterogeneity among players. Two classes of stochastic differential games, namely distributed games and games with mean field interactions, are analyzed to highlight the dependence of α on general game characteristics. To analyze the α-NE, the associated optimization problem is embedded into a conditional McKean-Vlasov control problem. A verification theorem is established to construct α-NE based on solutions to an infinite-dimensional Hamilton-Jacobi-Bellman equation, which is reduced to a system of ordinary differential equations for linear-quadratic (LQ) games. We conclude by case-studying an N-player LQ game on a graph network, analyzing α under different graph structures, and deriving the explicit solutions to the α-NE. Since our framework reduces multi-agent games to a single optimization problem, the second part of this thesis focuses on designing efficient algorithms for single-agent reinforcement learning (RL). While much progress has been made in RL for discrete Markov decision processes, continuous RL remains less explored. Therefore, in Chapter 4, we propose and analyze two new policy learning methods: regularized policy gradient (RPG) and iterative policy optimization (IPO), for a class of discounted linear-quadratic control (LQC) problems with continuous state space and continuous action space, over an infinite time horizon with entropy regularization. Assuming access to the exact policy evaluation, both proposed approaches are proved to converge linearly in finding optimal policies. Moreover, the IPO method can achieve a super-linear convergence rate once it enters a local region around the optimal policy. Finally, when the optimal policy for an RL problem in a known environment is appropriately transferred as the initial policy to an RL problem in an unknown environment, the IPO method is shown to converge at a super-linear rate if the two environments are sufficiently close. A model-free version of the policy-based methods is also discussed. Performances of these proposed algorithms are supported by numerical examples.
- 일반주제명
- Applied mathematics
- 일반주제명
- Statistics
- 일반주제명
- Mathematics
- 키워드
- Nash equilibria
- 키워드
- Markovian policy
- 기타저자
- University of California, Berkeley Industrial Engineering & Operations Research
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358464
■00520260202104706
■006m o d
■007cr#unu||||||||
■020 ▼a9798288866678
■035 ▼a(MiAaPQ)AAI32117150
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a519
■1001 ▼aLi, Xinyu.
■24513▼aAn α-Potential Game Framework for Non-Cooperative Dynamic Games: Theory and Algorithms
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a164 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Guo, Xin.
■5021 ▼aThesis (D.Eng.)--University of California, Berkeley, 2025.
■520 ▼aMulti-agent systems naturally arise in many real-world scenarios where multiple decision-makers interact, such as autonomous driving, network design, and financial markets. Analyzing multi-agent non-cooperative games is inherently challenging due to the asymmetry among players and the diverse structures of the games. This thesis introduces a unified framework to analyze N-player non-cooperative games in dynamic settings, considering both discrete-time and continuous-time state transitions. Additionally, efficient reinforcement learning (RL) algorithms and stochastic control techniques are proposed to identify strategies for each player that lead to approximate Nash equilibria (NE).The first part of this thesis introduces and analyzes a general class of dynamic N-player non-cooperative games called α-potential games. In this framework, the change in a player's objective function resulting from a unilateral deviation from her strategy is equal to the change in a common function, called the α-potential function, up to an error α. The existence of an α-potential function simplifies the challenging task of finding α-Nash equilibria in dynamic games to minimizing the α-potential function, as the optimizer of the α-potential function is shown to be an α-Nash equilibrium of the game.In Chapter 2, we focus on Markov games with finite state space, finite action space, and Markovian policy. The state transition follows a discrete-time Markov decision process. In this case, we establish the existence of an associated α-potential function. Additionally, we provide a semi-infinite linear program to find α and its corresponding α-potential function for any Markov game. We study two important classes of practically significant Markov games, Markov congestion games and the perturbed Markov team games, via the framework of Markov α-potential games, with explicit characterization of an upper bound for α and its relation to game parameters. Furthermore, we study two equilibrium approximation algorithms, namely the projected gradient-ascent algorithm and the sequential maximum improvement algorithm, along with their Nash regret analysis, and corroborate the results with numerical experiments using model-free RL algorithms.In Chapter 3, we study dynamic games with continuous-time state transitions, focusing on general classes of states, actions, and controls/policies, with a particular emphasis on stochastic differential games. An analytical characterization of the α-potential function is established, with α represented in terms of the magnitude of the asymmetry of the second-order derivatives of the players' objective functions. For stochastic differential games in which the state dynamic is a controlled diffusion, α is characterized in terms of the number of players, the choice of admissible strategies, and the intensity of interactions and the level of heterogeneity among players. Two classes of stochastic differential games, namely distributed games and games with mean field interactions, are analyzed to highlight the dependence of α on general game characteristics. To analyze the α-NE, the associated optimization problem is embedded into a conditional McKean-Vlasov control problem. A verification theorem is established to construct α-NE based on solutions to an infinite-dimensional Hamilton-Jacobi-Bellman equation, which is reduced to a system of ordinary differential equations for linear-quadratic (LQ) games. We conclude by case-studying an N-player LQ game on a graph network, analyzing α under different graph structures, and deriving the explicit solutions to the α-NE. Since our framework reduces multi-agent games to a single optimization problem, the second part of this thesis focuses on designing efficient algorithms for single-agent reinforcement learning (RL). While much progress has been made in RL for discrete Markov decision processes, continuous RL remains less explored. Therefore, in Chapter 4, we propose and analyze two new policy learning methods: regularized policy gradient (RPG) and iterative policy optimization (IPO), for a class of discounted linear-quadratic control (LQC) problems with continuous state space and continuous action space, over an infinite time horizon with entropy regularization. Assuming access to the exact policy evaluation, both proposed approaches are proved to converge linearly in finding optimal policies. Moreover, the IPO method can achieve a super-linear convergence rate once it enters a local region around the optimal policy. Finally, when the optimal policy for an RL problem in a known environment is appropriately transferred as the initial policy to an RL problem in an unknown environment, the IPO method is shown to converge at a super-linear rate if the two environments are sufficiently close. A model-free version of the policy-based methods is also discussed. Performances of these proposed algorithms are supported by numerical examples.
■590 ▼aSchool code: 0028.
■650 4▼aApplied mathematics
■650 4▼aStatistics
■650 4▼aMathematics
■653 ▼aMulti-agent systems
■653 ▼aNash equilibria
■653 ▼aReinforcement learning
■653 ▼aMarkovian policy
■653 ▼aStochastic control
■690 ▼a0364
■690 ▼a0800
■690 ▼a0405
■690 ▼a0463
■71020▼aUniversity of California, Berkeley▼bIndustrial Engineering & Operations Research.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0028
■791 ▼aD.Eng.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358464▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


