Relationship between different concepts always confuses me, but your video explained it in a explicit diagram and this really helpes me a lot. Feeling really thrilled.Thanks again!
i love the way you explain it through the formula's most experts tell you the formula then go to an actual case, which leaves the learner disconnected from the math, thanks!
I've watched other lectures on RL before, I can understand the formulas much better now, the way you explain formulas is brilliant, you're a wonderful math lecturer
Thank you so much. I've watched a lot of videos and didn't fully get these concepts for some reason. Now I think I finally get it. You're a great teacher.
The bellman’s equation reminds me of sequential games from game theory, where you traverse a game tree and optimal choice at each branch leads to globally optimal states.
At 3:57, I think the R(s', s, a) function you are referring to is the "reward function", which returns the "Immediately reward (r) if you are at stage (s) and do the action (a) which lead to stage (s')". That would make more sense than "returning a PROBABILITY of a reward (r) given (s, a and s')". I saw this in your book also but cannot find this kind of function anywhere else. All other resources I found, when talking about this function R, that means the "immediately reward" of doing action a given stage s and new stage s', NOT the "probability of the reward". Later on in the clip, when you uses it in value function, I also see you use it as a mean for measuring the "Value of reward", not the "Probability of reward", therefore I think this might really be a mistake or something. If I'm getting it wrong somewhere, please help me clear my thought. I'm just being curious. Love your great work.
Thank you for simplifying a lot of things. I had read corresponding chapters from Sutton and Barto book but I got more clarity on practical aspects from this video.
16:55 The value iteration function (VI) differs slightly from Bellman's equation (BE) because VI uses max on a (hence uses a single value), whereas BE uses max on all pi. Because pi is a probabilistic function, i.e. is yielding a specific action value 'a' with a certain probability, VI would need to have another level of summation over a multiplying the terms by pi(s,a). 20:05 Here we construct pi(s,a) as the argmax of VI. This means we set pi(s, argmax(s))=1, and pi(s, a')=0 for all other values a' /= argmax(s). This means pi(s,a) is deterministic, instead of probabilistic.
Beautiful. Please continue. Will you explain algorithms like PPO, TD3, DDPG, etc.? If so, I will appreciate each one. Also, it will be very interesting if you can give your opinion on some RL libraries like ray/RLlib, baselines3, etc. I know that this may be much more than what you are thinking of including in this course, but I do not lose anything by suggesting those topics to you :) Thank you.
Hi professor. Thanks for the wonderful videos. I was wondering why you classify actor critic as model based when the model of the environment is not learnt in this algorithm?
I love the way you give the material. Became curious about how do you project those formulae onto screen and able to see them? Is it glass screen and projector on side of camera? Or is it special screen?
Dear Dr. Steve I have a question I think according to what you explain to us, in value iteration we need to use an optimal algorithm; however, in policy iteration we don’t need to use that isn’t it Im looking forward to hearing from you Sincerely mohammad
Dear Dr.steve I have a question I think in value iteration we need to use an optimal algorithm;however, in policy iteration we don’t need to use that is it true?
I strongly recommend the book Reinforcement Learning an Introduction by Sutton and Barto. Also the Winter 2019 online lectures by Stanford (on TH-cam). Both have lots of maths and programming exercises.
thank you, but why does it always have to be so theoratical. Why cant you show an example like the tic tac toe which you mentioned to explain value iteration
Hi, Thank you for this great video. Can you please explain how we can use a model of system (for example LSTM) that predicts future as a simulator to run our reinforcement learning algorithm in it. So assume I trained a RL algorithm via model-free approach, but I can't test it on real environment and I need to test it on a simulated environment. How can we do this with having a model for prediction of the future via time-series data?
Interesting, funny that model based learning isn't highly regarded and so maybe not as explored. I get the feeling that this method will turn out to be as important as the data table function in excel. Quietly, and matter of factly determining a lot of our daily lives. The number of excel simulation models that impact our daily lives, is kinda scary. (think banks, insurance, etc back in the 90's and 2000's. think of all the mergers that were run through an "excel model", all the go/no go business decisions determined by excel models, all based on the data table simulation process, I'm sure model based deep learning has already taken over a lot of that, problem is no one wants to share their business secret sauce, and academia isn't interested in exploring this further. Shame.
Dear Dr. Steve I have a question I think according to what you explain to us, in value iteration we need to use an optimal algorithm; however, in policy iteration we don’t need to use that isn’t it Im looking forward to hearing from you Sincerely mohammad
Could you possibly explain more about "policy iteration and value iteration, leading to the quality function"? 25:40. Specifically, what is "redundant"? I believe there is a mistake. Here Q(s, a) and V_\pi(s) seem to have exactly the same formation. I still did not understand how this lead to the conclusion that quality function allows us to enable "model-free learning". I think the correct formula for Q is : Q(\mathbf{s}, \mathbf{a})=\mathbb{E}\left(R\left(\mathbf{s}^{\prime}, \mathbf{s}, \mathbf{a} ight)+\gamma Q\left(\mathbf{s}^{\prime}, a ight) ight) By the way, I am also a little confused about what is the "model" of the future reward is? 25:10
This is just awesome, especially for an undergraduate without much pre-knowledge about machine learning.Many thanks from a Chinese freshman.
Relationship between different concepts always confuses me, but your video explained it in a explicit diagram and this really helpes me a lot. Feeling really thrilled.Thanks again!
i love the way you explain it through the formula's most experts tell you the formula then go to an actual case, which leaves the learner disconnected from the math, thanks!
your videos are increadibly well thought out and very educational, I should have known about them sooner. greetings from Munich, Germany!
Thank you for making these videos. I'm learning so much! You are such a great explainer.
I've watched other lectures on RL before, I can understand the formulas much better now, the way you explain formulas is brilliant, you're a wonderful math lecturer
I actually feel smarter after watching this. Excellent video on all fronts!
Thank you so much. I've watched a lot of videos and didn't fully get these concepts for some reason. Now I think I finally get it. You're a great teacher.
Optimal control, Control Theory, Reinforcement Learning, Machine Learning, System Theory, System Identification are intellectual banquet.
probably the best explanation
Thank you Prof! this video really helpful to classify RL's methods. I really appreciate your diagram and your explanation.
Thanks -- glad it is helpful!
Love this series! Hoped the video to go on and on but it ended too quickly. Can't wait for the next part! Keep up the great work :)
Thanks so much!!
SUPERBBBBBB! Now I understand every inch of the research paper I was reading. Thanks!!!!
The bellman’s equation reminds me of sequential games from game theory, where you traverse a game tree and optimal choice at each branch leads to globally optimal states.
This is an excellent companion to your book. Thanks for both!
Very well structured and layed out, clearly explained, thank you
At 3:57, I think the R(s', s, a) function you are referring to is the "reward function", which returns the "Immediately reward (r) if you are at stage (s) and do the action (a) which lead to stage (s')". That would make more sense than "returning a PROBABILITY of a reward (r) given (s, a and s')". I saw this in your book also but cannot find this kind of function anywhere else. All other resources I found, when talking about this function R, that means the "immediately reward" of doing action a given stage s and new stage s', NOT the "probability of the reward".
Later on in the clip, when you uses it in value function, I also see you use it as a mean for measuring the "Value of reward", not the "Probability of reward", therefore I think this might really be a mistake or something.
If I'm getting it wrong somewhere, please help me clear my thought. I'm just being curious.
Love your great work.
Tnx a lot professor Brunton!
You're creating great materials!
Thank you for simplifying a lot of things. I had read corresponding chapters from Sutton and Barto book but I got more clarity on practical aspects from this video.
16:55 The value iteration function (VI) differs slightly from Bellman's equation (BE) because VI uses max on a (hence uses a single value), whereas BE uses max on all pi. Because pi is a probabilistic function, i.e. is yielding a specific action value 'a' with a certain probability, VI would need to have another level of summation over a multiplying the terms by pi(s,a).
20:05 Here we construct pi(s,a) as the argmax of VI. This means we set pi(s, argmax(s))=1, and pi(s, a')=0 for all other values a' /= argmax(s). This means pi(s,a) is deterministic, instead of probabilistic.
Beautiful. Please continue. Will you explain algorithms like PPO, TD3, DDPG, etc.? If so, I will appreciate each one. Also, it will be very interesting if you can give your opinion on some RL libraries like ray/RLlib, baselines3, etc. I know that this may be much more than what you are thinking of including in this course, but I do not lose anything by suggesting those topics to you :) Thank you.
Great suggestions! I will think about how to add these in the future. Might need to be in a future filming session, since it might take some time.
PPO would be very welcome. Deep RL is big now. Thanks for your videos, Dr. Long time fan
Good suggestion, hope we get videos on those soon
Thank you so much. Really appreciated the explanation at 24:20
now i know i ll understand well RL when you ll explain it!
Great and clear explanation, Steve! Thank you.
great video thank you for your contribution 🎉
Hi professor. Thanks for the wonderful videos. I was wondering why you classify actor critic as model based when the model of the environment is not learnt in this algorithm?
I love the way you give the material.
Became curious about how do you project those formulae onto screen and able to see them?
Is it glass screen and projector on side of camera? Or is it special screen?
Hi Steve, could you please add the videos to a playlist to avoid accidentally skipping videos?
Good call -- just added to playlist
thank you for the video, i wonder if there is a value function algorithm which is V(s,t)? Value of state s in time t
Thanks Professor Steve
Dear Dr. Steve I have a question
I think according to what you explain to us, in value iteration we need to use an optimal algorithm; however, in policy iteration we don’t need to use that isn’t it
Im looking forward to hearing from you
Sincerely mohammad
Amazing Content
Thank you
Recursion must be what supply's the reinforcement (feedback) to the value functions and eventually policy. Otherwise we're flying blind.
Dear Dr.steve I have a question
I think in value iteration we need to use an optimal algorithm;however, in policy iteration we don’t need to use that is it true?
Can i find some more detail? or a code with example of bellman equation??
I strongly recommend the book Reinforcement Learning an Introduction by Sutton and Barto. Also the Winter 2019 online lectures by Stanford (on TH-cam). Both have lots of maths and programming exercises.
@@RobinCarter thanks for your reply.
@@RobinCarter Agreed, these are great resources
Let me suggest the book:
Grokking Deep Reinforcement Learning by Miguel Morales
thank you, but why does it always have to be so theoratical. Why cant you show an example like the tic tac toe which you mentioned to explain value iteration
Hi, Thank you for this great video. Can you please explain how we can use a model of system (for example LSTM) that predicts future as a simulator to run our reinforcement learning algorithm in it. So assume I trained a RL algorithm via model-free approach, but I can't test it on real environment and I need to test it on a simulated environment. How can we do this with having a model for prediction of the future via time-series data?
WOOOOOOOOOO THERE IT IS, TIC TAC TOE HAS BEEN SOLVED
th-cam.com/video/xJR1oTDt1Ak/w-d-xo.html
How many agents can we have in the model?
implicit rolles theorem.
Interesting, funny that model based learning isn't highly regarded and so maybe not as explored. I get the feeling that this method will turn out to be as important as the data table function in excel. Quietly, and matter of factly determining a lot of our daily lives. The number of excel simulation models that impact our daily lives, is kinda scary. (think banks, insurance, etc back in the 90's and 2000's. think of all the mergers that were run through an "excel model", all the go/no go business decisions determined by excel models, all based on the data table simulation process, I'm sure model based deep learning has already taken over a lot of that, problem is no one wants to share their business secret sauce, and academia isn't interested in exploring this further. Shame.
wow, you talked 28 minutes and didn't solve any optimization problem with the techniques. I hope you know practically how to apply anything.
Dear Dr. Steve I have a question
I think according to what you explain to us, in value iteration we need to use an optimal algorithm; however, in policy iteration we don’t need to use that isn’t it
Im looking forward to hearing from you
Sincerely mohammad
Could you possibly explain more about "policy iteration and value iteration, leading to the quality function"? 25:40. Specifically, what is "redundant"?
I believe there is a mistake. Here Q(s, a) and V_\pi(s) seem to have exactly the same formation. I still did not understand how this lead to the conclusion that quality function allows us to enable "model-free learning".
I think the correct formula for Q is :
Q(\mathbf{s}, \mathbf{a})=\mathbb{E}\left(R\left(\mathbf{s}^{\prime}, \mathbf{s}, \mathbf{a}
ight)+\gamma Q\left(\mathbf{s}^{\prime}, a
ight)
ight)
By the way, I am also a little confused about what is the "model" of the future reward is? 25:10
Thanks for the lecture,Are value interation and policy interation learning aslo Temporal Difference learning?
In 8:20, shouldn't we have considered all the possible states s1=s' we can possibly end up when we follow a policy π?
I am too dumb to understand this.
Another great work. Thanks again.