An algorithm to determine the optimal control of a discrete system (Q2577208)

Based on the works done by Krotov, the authors introduced the iterative and improving algorithm for determining the optimal control of a discrete system. The article which was translated from the Russian, in an almost hard way, explains the sufficiently effective method for obtaining (at least) a local minimum of such a system. The main attempt of this method was to improve an element so that the given functional decreases and comes to the problem of the intermediary functional in which it contains a metric-like functional term. Thereafter, by regarding [\textit{L. N. Nikiforova} and \textit{M. Yu. Ukhin}, Approximate design of discrete optimal control, in Tr. Konf., Posvyashchennoi 20-letiyu IPS RAN (Proc. Conf. 20th Anniversary of the Program Systems Institute, Russian Academy of Sciences), Moscow: Fizmatlit, 377--386 (2004)], for a discrete model which is a function of its final state, based on the Taylor approximation of the Krotov-Bellman function over the grid of nodes, the algorithm was constructed. In this manner, according to Krotov's theory and using the generalized Lagrangian function, an approximated solution is determined by the use of the minimum conditions of a sum of the first and second variations of the Lagrangian. Indeed the transferred problem contains two controllers including a coefficient in \([0,1]\) and the variation of time interval. The aim of the time controller lies in making the set of time instants where the control is varied so that the resulting trajectory remains within the neighborhood of an initial point. However, it may be insufficient just to specify a required subset of the discrete time interval; hence the artificial points are defined and the system was redefined at these points. Afterwards, the algorithm is described for the three possible cases. It was also claimed that the algorithm is convergent of the second order and it is shown that it has more facilities for adaptation to a particular problem. At the end, two examples were given; in case one the penalty function is used to relax the constraints. In the first example which is methodological, an obtained maximum valve was taken as the initial approximation for the improvement procedures. In the second one, which is descriptive, the problem was solved in two steps. First, the turnpike solution was directly (i.e. independent of the boundary conditions) determined to use as an efficient initial approximation. Then, the mentioned first-order algorithm was applied.