Convergence Analysis of Evolutionary Algorithms That Are Based on the Paradigm of Information Geometry

被引:38
作者
Beyer, Hans-Georg [1 ]
机构
[1] Vorarlberg Univ Appl Sci, Dept Comp Sci, A-6850 Dornbirn, Austria
基金
奥地利科学基金会;
关键词
Convergence analysis; evolution strategies; information gain; information geometry; relative entropy loss; OPTIMIZATION; STRATEGY;
D O I
10.1162/EVCO_a_00132
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
The convergence behaviors of so-called natural evolution strategies (NES) and of the information-geometric optimization (IGO) approach are considered. After a review of the NES/IGO ideas, which are based on information geometry, the implications of this philosophy w.r.t. optimization dynamics are investigated considering the optimization performance on the class of positive quadratic objective functions (the ellipsoid model). Exact differential equations describing the approach to the optimizer are derived and solved. It is rigorously shown that the original NES philosophy optimizing the expected value of the objective functions leads to very slow (i.e., sublinear) convergence toward the optimizer. This is the real reason why state of the art implementations of IGO algorithms optimize the expected value of transformed objective functions, for example, by utility functions based on ranking. It is shown that these utility functions are localized fitness functions that change during the IGO flow. The governing differential equations describing this flow are derived. In the case of convergence, the solutions to these equations exhibit an exponentially fast approach to the optimizer (i.e., linear convergence order). Furthermore, it is proven that the IGO philosophy leads to an adaptation of the covariance matrix that equals in the asymptotic limit-up to a scalar factor-the inverse of the Hessian of the objective function considered.
引用
收藏
页码:679 / 709
页数:31
相关论文
共 28 条
  • [1] Akimoto Youhei, 2012, Parallel Problem Solving from Nature - PPSN XII. Proceedings of the 12th International Conference, P42, DOI 10.1007/978-3-642-32937-1_5
  • [2] Analysis of a Natural Gradient Algorithm on Monotonic Convex-Quadratic-Composite Functions
    Akimoto, Youhei
    [J]. PROCEEDINGS OF THE FOURTEENTH INTERNATIONAL CONFERENCE ON GENETIC AND EVOLUTIONARY COMPUTATION CONFERENCE, 2012, : 1293 - 1300
  • [3] Theoretical Foundation for CMA-ES from Information Geometry Perspective
    Akimoto, Youhei
    Nagata, Yuichi
    Ono, Isao
    Kobayashi, Shigenobu
    [J]. ALGORITHMICA, 2012, 64 (04) : 698 - 716
  • [4] Natural gradient works efficiently in learning
    Amari, S
    [J]. NEURAL COMPUTATION, 1998, 10 (02) : 251 - 276
  • [5] [Anonymous], 2010, P 12 ANN C GENETIC E
  • [6] [Anonymous], 1993, ESIMATION THEORY
  • [7] Arnold B.C., 1998, RECORDS
  • [8] Evolutionary gradient search revisited
    Arnold, Dirk V.
    Salomon, Ralf
    [J]. IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, 2007, 11 (04) : 480 - 495
  • [9] Weighted multirecombination evolution strategies
    Arnold, Dirk V.
    [J]. THEORETICAL COMPUTER SCIENCE, 2006, 361 (01) : 18 - 37
  • [10] Performance analysis of evolutionary optimization with cumulative step length adaptation
    Arnold, DV
    Beyer, HG
    [J]. IEEE TRANSACTIONS ON AUTOMATIC CONTROL, 2004, 49 (04) : 617 - 622