We have observed one of the serious problems with on-line parameter identification used for flight control: the ill-conditioning problem that can result in the estimated parameters blowing up when the airplane is cruising. The sequential least squares algorithm with a forgetting factor is higly susceptible to this.
We saw two remedies for this problem. The first was to increase or eliminate the forgetting factor. This resulted in acceptible controller performance for the maneuver attempted. However, the lack of a forgetting factor was very detrimental to the accurate estimation of the parameters.
The second remedy was to decrese the least squares update rate. This resulted in an estimator that worked, in that it reasonably estimated the parameters after they had changed. However, its robustness is to be doubted, for it seems to depend on a certain amount of noise for its stability.
Finally, we saw a possible hint of another problem: parameters that are poorly estimated because not enough data is collected in condition in which they domination.