PLANNING MACHINE EXPERIMENTS WITH SYSTEM MODELS
Simulation modeling is inherently a computer experiment with a model, an explored or projected system. The plan of an imitation experiment on a computer is a method of obtaining, through the experiment, information that the user needs. The effectiveness of using experimental resources depends significantly on the choice of the experimental design. The main goal of experimental research using simulation models is to study the behavior of the simulated system in the most profound way. For this, it is necessary to plan and design not only the model itself, but also the process of its use, that is, carrying out experiments with it on a computer. The whole complex of questions of planning experiments with simulation models for their successful solution can be rationally divided into strategic and tactical planning.
METHODS OF THE THEORY OF PLANNING EXPERIMENTS
A computer experiment with a model of system 5 during its investigation and design is conducted with the aim of obtaining information on the characteristics of the process of functioning of the object under consideration. This information can be obtained both for the analysis of characteristics and for their optimization under specified constraints, ie, for synthesizing the structure, algorithms, and parameters of the system. Depending on the goals of modeling the system 5 on a computer, there are different approaches to organizing the simulation experiment with the machine model M and . The main task of planning machine experiments is obtaining the necessary information about the system under study 5 with resource constraints (computer time, memory, etc.). Particular problems solved in the planning of machine experiments include the tasks of reducing the expenditure of computer time for modeling, increasing the accuracy and reliability of the results of modeling, checking the adequacy of the model, etc.
The effectiveness of computer experiments with the models M m essentially depends on the choice of the experimental design, since the plan determines the amount and order of computations on computers, the methods of accumulation and statistical processing of results Therefore, the main task of planning machine experiments with the model M and is formulated as follows: it is necessary to obtain information about the modeling object specified in the form of a modeling algorithm (program), with minimal or limited costs x machine resources for the implementation of the modeling process.
Thus, in computer modeling, it is rational to plan and design not only the model M s of system 5, but also the process of its use, ie, carrying out experiments with it using an instrumental computer.
To date, the theory of experimental planning has developed in physics, biology, etc., in which sufficiently powerful mathematical methods have been developed that make it possible to increase the efficiency of such experiments [10, 18, 21, 33]. But transfer of these results to the field of computer experiments with models M and can take place only taking into account the specificity of computer systems simulation. Despite the fact that the goals of experimental computer simulations and actual experiments coincide, there are differences between these two types of experiments, therefore the most important for planning the experiment is the following: 1) the simplicity of repeating the experimental conditions on a computer with the model M s systems 5; 2) the ability to control the experiment with the model L/ m , including its interruption and resumption; 3) ease of variation of the experimental conditions (environmental influences & pound;); 4) the presence of a correlation between the sequence of points in the modeling process; 5) difficulties associated with the definition of the simulation interval (0, 7).
The advantage of computer experiments before the actual is the ability to fully reproduce the experimental conditions with the model of the system being analyzed. It is possible to compare two alternatives under the same conditions, which is achieved, for example, by choosing the same sequence of random numbers for each of the alternatives. A significant advantage over full-scale is the ease of interrupting and resuming machine experiments, which makes it possible to apply sequential and heuristic planning techniques that may not be feasible in experiments with real objects. When working with the machine model M m , it is always possible to interrupt the experiment for the time necessary to analyze the results and make decisions about its further progress (for example, > M m ).
The disadvantage of computer experiments is that often there are difficulties associated with the presence of correlation in the output sequences, that is, the results of some observations depend on the results of one or more previous ones, and therefore
they contain less information, it in independent observations. Since most of the existing methods for planning experiments assume the independence of observations, many of these methods can not be directly applied to computer experiments in the presence of correlation [18, 21, 29, 46, 53].