Been proposed for on-line identification of the states and parame ters of linear discrete-time multivariable systems. It is a one-stage algorithm, in which a special canonical form of the state equations provides a pseudo-linear rep resentation where ali unknown parameters and state variables appear linearly multi plied by either external variables, or by matrices that are only composed of zeroes and ones. Deriving an augmented system, that involves augmenting the state variables of the system by adjoining to them the unknown parameter vectors and treating them as a part of the new state variable vector, unbiased estimates of the parameters are ob tained using a discrete-time Kalman filter. The algorithm requires gaussian noise sequences with known statistics. A nice feature of the given algorithm is that, it works directly on input/output data and leads to a state representation. On the basis of the number of simulated examples treated. the identification approach appears to be working quite weIl. The convergence of the algorithm is very fast, the number of iterations needed to obtain conclusive convergence is very small, yet the estimation accuracy compares favorably with the performance of other more elaborate identification techniques. Actual/y, the minimal number of rterations nec essary for convergence is equal ta the total number of states and parameters ta be estimated. By reducing the number of iterations, the method requires legs computer time. It gives unbiaged and consistent estimates even for high noise-to-signal ratios. As ta be expected, the examples considered illustrate the importance of using an input sufficiently rich in spectral components. Tao, it is extremeJy important ta use initial conditions on the augmented state and on the covariance matrices which are nat in contradiction with the actual values of the parameters ta be identified. The identification method described here constitutes an effective on-line estimation technique which may be of value in practica/ situations. Though the preliminary re sults are encouraging, further work is needed ta make the algorithm operational in a real time fashion and speed up the computational routines in the program as weil as compare its performance in relation with other identification schemes. .In order ta imp/ement this estimation algorithm it is necessary ta know the spectral components, which represent our knowledge about the erturbations and noises. Ta estimate the noise characteristics can be advantageously used in algorithms of the type [ J, where an on-line a/gorithm for the identification of the model is proposed, first the autocovariance matrices are estimated and then a minimal realization is de rived. An investigation can be performed ta check whether the KaJman filter is working optimally or nat. If the Ka/man fiJter is suboptima/, a technique as mentioned above can be developed ta obtain asymptotically normal, unbiased and consistent estimates of Q and R. Also, the problem of structure estimation for the state space mode' from noisy data is of great importance in system identification and more work shouJd be done about it . From practical considerations, it is important to determine the adequacy of the model under the types of inputs the process is likely to encounter.