11.1. In the previous chapter, various minimax problems were investigated, where either the maximum or minimum were to be computed over a finite-dimensional set. Several ways of transforming these problems to problem 5 were suggested (if the exterior minimum is finite-dimensional, then the transformation is carried out in the same way as at the end of Subsec.10.5.). In the general case of infinite-dimensional minimax problems without the assumption of concavity w.r.t, a maximizing variable, one discovers, following the recommended scheme, that an integral appears w.r.t, a measure defined on a set of an infiniteZdimensional space. The question of the finite-dimensionM approximation with regard to such Optimization problems involving integrals has not yet been properly worked out, and the more So, the methods of search for the optimum. Thus the possibility of constructing simple, though not highly precise, numerical algorithms for the solution of the problem seems to be somewhat important. Such algorithms can be elaborated within CMIPAS. Below we shall demonstrate this possibility, using a generalization of problem 1 (Sec. 1) for the infinite-dimensional case as an example. Further development of this example (in Subsec. 2) leads to a minimax problem whose solution may also be obtained within CMIPAS. We prove the convergence of the corresponding algorithm in Sec. 3. The suggested algorithm is applicable to any minimax problem in L2~ which is convex w.r.t, a minimizing variable. Methods for solving convex-concave minimax problems (search for a saddle point) in Hilbert space will be considered in the next section (Subsecs. 12.3-5).Problem 11 (a generalization of problem 1). Find the value I ° and the solution u ° E U' of the infimum:inf I(u),u 6 U V where u = { . e u ' c r . ( X ) l a ( . ) ( x ) ___ 0 a . s . V , e X},
V = {v E V' C L2(Y)IH(v)(y) < 0 a.s. Vy E Y};V is #-measurable; U', V' are weakly compact sets; g(.)(y) and G(.)(z) are continuous functionals on Y' and U', respectively, £a.a. y 6 Y C R m, x E X C Rm'; J(u, .) are #-integrable VuE U'. Suppose that the functional J can be represented as follows:where ~(.,-,-) is a function of three variables; g(-), h(.), and f(.,-) are operators from U', V', and U' x V' into L2, respectively. The measure # is a probability measure defined on V'. When Vu 6 U' J(u, .) is integrable w.r.t, a finite-additive measure (a weak distribution [121]), we do not require the measure # to be a-additive. If # is finite-additive, then we assume that 3 c < + o c : J ( . , . ) > -c .In stochastic optimization models, the measure # is not available in explicit form, but one can observe the corresponding independent values v 6 V'. However, such an observation seems to be problematic in the case of an infinite-dimensional space. Therefore, let us consider this possibility with regard to only finitedimensional projections of the measure #, and let us use (in the algorithm) independent values of random variables, distributed in accordance with the measures #n defined on R n, n = 1, 2, .... ...