On Parallel Complexity of Nonsmooth Convex Optimization
暂无分享,去创建一个
We consider the standard class of problems ?(x) ? min, x ? Bn associated with convex continuous functions ? mapping the unit n-dimensional cube Bn into 0, 1]. It is known that the information complexity of the class with respect to the standard first-order oracle is, within an absolute constant factor, n ln (1/?), ? < 12 being the required accuracy (measured in terms of ?). The question we are interested in is how the complexity can be reduced if one is allowed to use K copies of the oracle in parallel rather than a single oracle. We demonstrate that the "K-oracle complexity" is at least O(1)(n/ln(2Kn))1/3ln(1/?).
[1] Boleslaw Z. Kacewicz. On sequential and parallel solution of initial value problems , 1990, J. Complex..
[2] John Darzentas,et al. Problem Complexity and Method Efficiency in Optimization , 1983 .
[3] Stefan Heinrich,et al. Parallel information-based complexity , 1991, J. Complex..
[4] H. Woxniakowski. Information-Based Complexity , 1988 .
[5] Joachim von zur Gathen,et al. Parallel Arithmetic Computations: A Survey , 1986, MFCS.