Hi @brian, requesting you to make a series on black box and white box modelling technique dedicated to it with indepth analysis. This series is also really good....!
Thank you for the amazing contents, MATLAB providing, however, I request to the MATLAB community to make such a wonderful course on Signal Processing as well. Thank you
It can be noted that the recursive least squares (with exponential forgetting factor) is essentially a simplified version of a Kalman filter. And your statement about that the RLS with forgetting factor converges faster I believe isn't true, since one can probably also obtain the same behavior with a Kalman filter by using different covariance matrices. It is also worth nothing that these filters require a persistence of excitation condition. Namely, in that DC motor example it is not possible to estimate the friction if the motor is commanded to stand still.
No, you can not converge faster than Kalman / true RLS because they are BLUE. Full stop. No, finite memory LS can not outperform RLS. You can mess up their initialization, however - but that is a different issue. You may have mentioned multi-model approaches (instead of hopeless gain balancing) and Fredrikson's book.
kinda does, RL can be used to change gains of your controller of the system or identify parameters of your system (some RL algorithms use Neural Networks that are ell known for pattern and parameters recognition)
System Identification is a field in which Reinforcement Learning is one of the many methods that can be used to achieve system ID objective. Other popular method for system ID is neural network, linear regression, and non linear regression.
Is this used in finance? Anybody know? Or the behaviour of asset prices is solely random and has no systematic component? Hence, system identification won't work...because the stock market has no system in the first place?
If we are using measurement y in the regressor, what are we predicting? It is said yhat is prediction at time t that is equal to theta1*y(t-1) + theta2*u(t-1). If y(t-1) is measurement at previous time t-1, then it seems to be a filter rather than a predictor. Instead if yhat(t-1) is used in the regressor it appears as the predictor that predicts y as a function of u only. Isn’t that the reason for building a model to predict something as a function of independent variable.
I believe y(t) there is a typo. Only u(t) for the current time step makes sense. Also, a filter basically gives us the a posteriori state values given the a priori values predicted from the previous estimates states (and the model) and the current measurement. The predictor predicts one step forward given the measurement up to y(t-1) and input up to u(t) while the filter refines the state values given a priori prediction and current measurement. Hope this helps.
Thank you so much for your videos, Brian. Yours are the only reason I’m subscribed to Matlab’s channel!
Hi @brian, requesting you to make a series on black box and white box modelling technique dedicated to it with indepth analysis. This series is also really good....!
Thank you for the amazing contents, MATLAB providing, however, I request to the MATLAB community to make such a wonderful course on Signal Processing as well. Thank you
These are incredibly helpful
It can be noted that the recursive least squares (with exponential forgetting factor) is essentially a simplified version of a Kalman filter. And your statement about that the RLS with forgetting factor converges faster I believe isn't true, since one can probably also obtain the same behavior with a Kalman filter by using different covariance matrices.
It is also worth nothing that these filters require a persistence of excitation condition. Namely, in that DC motor example it is not possible to estimate the friction if the motor is commanded to stand still.
Awesome! Please keep these coming
No, you can not converge faster than Kalman / true RLS because they are BLUE. Full stop.
No, finite memory LS can not outperform RLS.
You can mess up their initialization, however - but that is a different issue.
You may have mentioned multi-model approaches (instead of hopeless gain balancing) and Fredrikson's book.
you broke ma brain
it seems like system id is quite related to reinforcement learning isnt it? model based RL specifically
kinda does, RL can be used to change gains of your controller of the system or identify parameters of your system (some RL algorithms use Neural Networks that are ell known for pattern and parameters recognition)
System Identification is a field in which Reinforcement Learning is one of the many methods that can be used to achieve system ID objective. Other popular method for system ID is neural network, linear regression, and non linear regression.
Is this used in finance? Anybody know? Or the behaviour of asset prices is solely random and has no systematic component? Hence, system identification won't work...because the stock market has no system in the first place?
If we are using measurement y in the regressor, what are we predicting? It is said yhat is prediction at time t that is equal to theta1*y(t-1) + theta2*u(t-1). If y(t-1) is measurement at previous time t-1, then it seems to be a filter rather than a predictor. Instead if yhat(t-1) is used in the regressor it appears as the predictor that predicts y as a function of u only. Isn’t that the reason for building a model to predict something as a function of independent variable.
I believe y(t) there is a typo. Only u(t) for the current time step makes sense. Also, a filter basically gives us the a posteriori state values given the a priori values predicted from the previous estimates states (and the model) and the current measurement. The predictor predicts one step forward given the measurement up to y(t-1) and input up to u(t) while the filter refines the state values given a priori prediction and current measurement. Hope this helps.