วีดีโอ

What of we took H=1 do we compute a lot of inuts so we have a smooth control?

Awesome video sir! Thank you!

The claim that the separation principle does not hold for POMDPs (15:00 - 16:00 in the video) is false. The belief state (conditional distribution of the present state based on the history of measurements) is a sufficient statistic. As pointed out in the video, you can update the belief in a recursive fashion (a generalized version of the Kalman filter). The principle of optimality (Bellman equation) still holds and it's possible to show that you can write it in terms of the belief state rather than the full measurement history. Consequently, the optimal action is a function of the belief. This is not an approximation. The separation principle holds for POMDPs. What does not hold in the POMDP case is certainty equivalence. The optimal POMDP policy will in general be different from the optimal MDP policy applied to the MMSE estimate of the state (the expected value of the belief). It's only in the LQG case that this works. The reason people say "POMDPs are hard" has nothing to do with the separation principle. There are two reasons: 1) Doing belief propagation in continuous state/action/measurement space is intractable. You would have to store the belief as a conditional distribution, which is easy to do in the LQG case where you know the belief will be Gaussian, but in general you would have to store some function that does not have a closed-form expression. Likewise, you can do dynamic programming, but again you have to store the value function, which is easy in the LQG case where you know it will be quadratic, but is difficult in general. There are various approaches for approximating the belief propagation and dynamic programming steps, which is what is done in practice to overcome these difficulties. 2) You can find the exact POMDP policy if the state, action, and observation spaces are finite and discrete. Then, your belief and optimal policy can be stored as matrices and implemented as lookup tables. Unfortunately, this is only practically useful for very modest POMDPs due to the curse of dimensionality.

Starts at 3:30

hi , what is the difference between UC and Day- ahead? please

Great Work Sir Thanks Lot of Sir

thank you very much!!!, this will be helpful for my future master degree thesis in MPC

Thanks for the lecture. Please you have code MatLab or Simulink file examples

BORING!

Power flow begins 31:26

Is there a set of notes? Thanks

You are favourite

Great lecture. I have one persisting question though. How do you understand a model with many free parameters?

Very illuminating lecture. You put things in very nice mathematical ideas like convex functions and have made an important point about the linear and non-linear regression using features.

How did you plot those cool rotating 3D scatter plots?

I tried to convince myself that if f(x) = Ax then the Jacobian of f(x) is just A 59:00 by using the first example. I did it out on paper and it didn't work out so I tried it using MATLAB and got the same answer. Take a look imgur.com/a/Nlyf1BZ please let me know what I did wrong. I found an A matrix that works but it is not the Jacobian of f(x)… imgur.com/a/aDt668A

your lecture videos are great by the way :) thanks for sharing

please suggest a book you are following

32:55 How is fixing a function actually better than erasing it entirely? They both deliver us globally non optimal solution

for the quiz, machine learning because the robot has to be capable to learn the preferred coffee quantity for each one in the office, I like a big and full cup, no sugar. good robot. thanks Zico.

convex optimization at 50:20

Matrix Calculus start at 48:50

REALLY HELPFUL

Nice! Can you comment on how you made the 3D animated (rotating) graph?

This was really helpful.

Greetings, are there any plans to upload the Spring 2016 iteration of this course please, as you cover Deep Learning in it. Thank you for sharing these lecture video

Thank you! very helpful

Hello, Zico, I'm working on robot scheduling problem for automated wet station in semiconductor process. I want to learn how to use MILP concept to solve this problem, I've known all the constraints, but I don't know how to start with the algorithm. So how can I start to learn the MILP codeing for the very beginning?

The instruction is great Can't wait to watch more videos about control from this instructor Great job

Hey, quick question about the objective function for the Sudoku. Are we actually optimizing with respect to this objective function or is it just used to "sort" the branches? If we are optimizing with respect to it, how do we do it? (As the function isn't linear).

Hello Mr.Kolter, Could you please advice me for good book talk about power flow?

Thanks. In many ways an excellent video, however in the beginning you talk of the magnet as if it has one pole. Later at about 43 min you in fact refer to a "one pole" magnet! AC is produced because when the field from the North pole cuts the conductor, current flows in one direction and in the opposite when the South pole comes around (load connected). The min number of poles in the generator is 2 and it must rotate at 3600 rpm for 60 Hz. You mention this as an unreasonable speed, but in Europe the vast majority of thermal sets run at 3000 rpm and I expect many thermal set in the US run at 3600. Hydro sets are typically multi-pole and so run at much slower speeds. However the stators are wound so that the rotating field of the rotor is still in synch with the rotating field produced by the current flowing in the stator coils. Thanks for the info regarding how the generator voltage must be increased to deliver reactive energy to inductive loads, very useful explanation.

Thank you !!!! Really need this for my thesis!

I am eager to know what program you are using to view the waveform in your computer.?

Thanks for the upload. I was reading up on kernel methods but struggling to understand parts of it, but this really helped clear a lot of my doubts regarding the topic.

it starts at @54:51. Great teaching.

the teacher is handsome and hot

Someone please give me the Power Presentation link.

Thanks for the lecture. It is not true, however, that MDPs only work for finite actions and states. Have a look at policy search RL methods, and stuff like Fitted Q-Iteration.

Hi, can you help me, I have fitted a non linear model, but now I need to make a test of significance for every parameter of my non linear model, How should I proceed??, I estimate my parameters with non linear least squares

very useful thanks.

brutta linguaccia!!!!!!!!

I got a head-start into Computer Vision through the last two lectures.Very intuitive way of teaching and definitely not sleep-inducing. Thanks a lot for posting these!

Nice lecture.

if there anyone listening u talking, they probable all dead,u kill them

Lecture begins at 5:10.

Great Record. It helps me a lot to get the basic idea of non-linear regression. Thx