The article presents the main most noteworthy research results on the identification problem and methods that are widely used in solving identification problems. The most significant here is the problem of modeling especially complex systems, for which it is impossible, even in general, but in constructive form for identification, to write down a class of models in which the system under study has the most accurate description in which there is one of them adequate to the available data obtained from experiments. If the available knowledge on the object under study does not allow us to write down a suitable class of models in mathematical form, then asymptotic modeling should be used, as is done in computational mathematics. Instead of differential, integral and other equations, for example, algebraic equations approximating them are solved, the solutions of which approach the exact ones with increasing dimension and discretization step. Various forms of asymptotic expansions are used in identification, including rational approximation as applied to irrational and infinite-dimensional systems. However, in the presence of errors in the data, the use of asymptotic modeling leads to the fact that, with intentions to increase the accuracy of the resulting model due to its large dimension, in many cases the identification problems becomes incorrectly posed. Therefore, we have to look for a compromise or trade-off between bias and variance, which should determine the best-quality model. In addition to the modeling problem, the article discusses two concepts on the basis of which most existing methods for solving identification problems are based. Both of them give their own interpretation of the errors present in the data, namely, stochastic and nonstochastic. With the stochastic concept, errors are treated as a random process or sequence. The concept of estimation consistency is introduced, which makes it possible to use widely the theory of statistics in justifying methods for solving identification problems. The nonstochastic identification paradigm allows for arbitrary but bounded uncertainties in the data, i.e. belonging to some bounded sets. The largest number of different methods, as well as algorithms and software implementations, are made within the framework of the stochastic concept. Within the framework of nonstochastic identification, we mainly consider various modifications of the so-called subspace methods. In relation to complex systems, problems were noted that did not allow the development of sufficiently universal and effective methods for solving identification problems. A number of research areas have been pointed out that need further development in order to obtain more accurate models of complex systems.