STOCHASTIC SADDLE POINT PROBLEMS WITH DECISION-DEPENDENT DISTRIBUTIONS

被引:6
|
作者
Wood, Killian [1 ]
Dall'Anese, Emiliano [1 ,2 ]
机构
[1] Univ Colorado Boulder, Dept Appl Math, Boulder, CO 80309 USA
[2] Univ Colorado Boulder, Dept Elect Comp & Energy Engn, Boulder, CO USA
基金
美国国家科学基金会;
关键词
saddle point problems; minimax problems; stochastic gradient; decision-dependent distributions; OPTIMIZATION; ALGORITHMS; COMMUNICATION; GRADIENT;
D O I
10.1137/22M1488077
中图分类号
O29 [应用数学];
学科分类号
070104 ;
摘要
This paper focuses on stochastic saddle point problems with decision-dependent distributions. These are problems whose objective is the expected value of a stochastic payoff function and whose data distribution drifts in response to decision variables-a phenomenon represented by a distributional map. A common approach to accommodating distributional shift is to retrain optimal decisions once a new distribution is revealed, or repeated retraining. We introduce the notion of equilibrium points, which are the fixed points of this repeated retraining procedure, and provide sufficient conditions for their existence and uniqueness. To find equilibrium points, we develop deterministic and stochastic primal-dual algorithms and demonstrate their convergence with constant step size in the former and polynomial decay step-size schedule in the latter. By modeling errors emerging from a stochastic gradient estimator as sub-Weibull random variables, we provide error bounds in expectation and in high probability that hold for each iteration. Without additional knowledge of the distributional map, computing saddle points is intractable. Thus we propose a condition on the distributional map-which we call opposing mixture dominance-that ensures that the objective is strongly-convex-strongly-concave. Finally, we demonstrate that derivative-free algorithms with a single function evaluation are capable of approximating saddle points.
引用
收藏
页码:1943 / 1967
页数:25
相关论文
共 50 条
  • [1] Online Saddle Point Tracking with Decision-Dependent Data
    Wood, Killian
    Dall'Anese, Emiliano
    LEARNING FOR DYNAMICS AND CONTROL CONFERENCE, VOL 211, 2023, 211
  • [2] Stochastic Optimization with Decision-Dependent Distributions
    Drusvyatskiy, Dmitriy
    Xiao, Lin
    MATHEMATICS OF OPERATIONS RESEARCH, 2023, 48 (02) : 954 - 998
  • [3] Stochastic Approximation with Decision-Dependent Distributions: Asymptotic Normality and Optimality
    Cutler, Joshua
    Diaz, Mateo
    Drusvyatskiy, Dmitriy
    JOURNAL OF MACHINE LEARNING RESEARCH, 2024, 25
  • [4] Online Projected Gradient Descent for Stochastic Optimization With Decision-Dependent Distributions
    Wood, Killian
    Bianchin, Gianluca
    Dall'Anese, Emiliano
    IEEE CONTROL SYSTEMS LETTERS, 2022, 6 : 1646 - 1651
  • [5] Distributionally robust facility location problem under decision-dependent stochastic demand
    Basciftci, Beste
    Ahmed, Shabbir
    Shen, Siqian
    EUROPEAN JOURNAL OF OPERATIONAL RESEARCH, 2021, 292 (02) : 548 - 561
  • [6] Models and applications of stochastic programming with decision-dependent uncertainty in power systems: A review
    Yin, Wenqian
    Hou, Yunhe
    IET RENEWABLE POWER GENERATION, 2024, 18 (14) : 2819 - 2834
  • [7] A decision-dependent randomness stochastic program for asset-liability management model with a pricing decision
    Kopa, Milos
    Rusy, Tomas
    ANNALS OF OPERATIONS RESEARCH, 2021, 299 (1-2) : 241 - 271
  • [8] A Decision-dependent Stochastic Approach for Wind Farm Maintenance Scheduling Considering Wake Effect
    Yin, Wenqian
    Peng, Xiaosheng
    Hou, Yunhe
    2020 IEEE PES INNOVATIVE SMART GRID TECHNOLOGIES EUROPE (ISGT-EUROPE 2020): SMART GRIDS: KEY ENABLERS OF A GREEN POWER SYSTEM, 2020, : 814 - 818
  • [9] Solving Decision-Dependent Games by Learning From Feedback
    Wood, Killian
    Zamzam, Ahmed S.
    Dall'Anese, Emiliano
    IEEE OPEN JOURNAL OF CONTROL SYSTEMS, 2024, 3 : 295 - 309
  • [10] Adaptive Stochastic Primal-Dual Coordinate Descent for Separable Saddle Point Problems
    Zhu, Zhanxing
    Storkey, Amos J.
    MACHINE LEARNING AND KNOWLEDGE DISCOVERY IN DATABASES, ECML PKDD 2015, PT I, 2015, 9284 : 645 - 658