Issues related to solving the additive problem of sequential traversal of sets with precedence restrictions and cost functions that allow dependence on the list of tasks are considered. The basic method is a broadly understood dynamic programming (DP), supplemented in the case of problems of appreciable dimension by decompositions of the family of tasks and transformation of the parameters of the original problem. Possible applications are related, in particular, to the problem of tool control in figured sheet cutting of parts on CNC machines. In this problem, an important circumstance is taking into account the precedence conditions, which have, in particular, the following meaning: in the case of a part with holes, cutting of each of the internal contours (corresponding to the holes) should precede cutting of the external contour. The quality criterion itself in this problem, as a rule, is additive. Another type of constraints concerns avoiding thermal deformations of parts. When using the approach with penalties for violating the conditions associated with effective heat dissipation during cutting, cost functions arise that allow dependence on the list of tasks completed to date. Note that in another applied problem, namely, in the problem of dismantling radiation hazardous objects, cost functions arise with dependence on the list of tasks that have not been completed at the moment (and, consequently, concern the objects that have not been dismantled). As a result, we arrive at a very general problem with precedence constraints and cost functions with dependence on the list of tasks. The decomposition applied in the case of a noticeable dimensionality with subsequent implementation of the DP requires, on the one hand, the development of clustering methods, and, on the other, the construction of an adequate structure for distributing global precedence conditions among clusters. In the theoretical part of the work, the case of two clusters is discussed, which makes it possible to cover with a single scheme a number of practically interesting problems of a range (in terms of dimensionality) type. An algorithm for constructing a composite solution is indicated, including a stage of clustering training based on a greedy algorithm. This “composite” algorithm is implemented on a PC; a computational experiment was carried out.
Read full abstract