In this course, you will learn how to solve problems with large, high-dimensional, and potentially infinite state spaces. You will see that estimating value functions can be cast as a supervised learning problem---function approximation---allowing you to build agents that carefully balance generalization and discrimination in order to maximize reward. We will begin this journey by investigating how our policy evaluation or prediction methods like Monte Carlo and TD can be extended to the function approximation setting. You will learn about feature construction techniques for RL, and representation learning via neural networks and backprop. We conclude this course with a deep-dive into policy gradient methods; a way to learn policies directly without learning a value function. In this course you will solve two continuous-state control tasks and investigate the benefits of policy gradient methods in a continuous-action environment. Prerequisites: This course strongly builds on the fundamentals of Courses 1 and 2, and learners should have completed these before starting this course. Learners should also be comfortable with probabilities & expectations, basic linear algebra, basic calculus, Python 3.0 (at least 1 year), and implementing algorithms from pseudocode. By the end of this course, you will be able to: -Understand how to use supervised learning approaches to approximate value functions -Understand objectives for prediction (value estimation) under function approximation -Implement TD with function approximation (state aggregation), on an environment with an infinite state space (continuous state space) -Understand fixed basis and neural network approaches to feature construction -Implement TD with neural network function approximation in a continuous state environment -Understand new difficulties in exploration when moving to function approximation -Contrast discounted problem formulations for control versus an average reward problem formulation -Implement expected Sarsa and Q-learning with function approximation on a continuous state control task -Understand objectives for directly estimating policies (policy gradient objectives) -Implement a policy gradient method (called Actor-Critic) on a discrete state environment
Great course. Slightly more complex than courses 1 and 2, but a huge improvement in terms of applicability to real-world situations.
more detailed explanation of some of the assignments and how state values are got with tile coding but overall a great experience!
I loved the course videos and programming assignments. The only suggestion would be to go a little deeper in the videos.
There is a lot of material covered in the course. Be aware the pace picks up considerably from the first two courses. This said, it is a worthwhile course to take.
This is the third instalment in reinforcement learning.so far so good. yeah, you can get stuck some times but it is okay you can make it out.
this course bridged the gap to Deep Learning, the most exciting direction in RL. I would like a sequel dedicated to this from U Alberta
Surely a level-up from the previous courses. This course adds to and extends what has been learned in courses 1 & 2 to a greater sphere of real-world problems. Great job Prof. Adam and Martha!
Solid intro course. Wish we covered more using neural nets. The neural net equations used very non-standard notation. Wish the assignments were a little more creative. Too much grid world.
Great content, Great presentation. I really appreciate the efforts that were made to create this comprehensive course.Thank you.
Martha and Adam are excellent instructors. This course is so well organized and presented. I have learned a lot! Thanks very much!
Well peaced and thoughtfully explained course. Highly recommended for anyone willing to set solid grounding in Reinforcement Learning. Thank you Coursera and Univ. of Alberta for the masterclass.
Great Learning, the best part was the Actor-Critic algorithm for a small pendulum swing task all from stratch using RLGLue library. Love to learn how experimentation in RL works.
Showing: 20 of 149
Fantastic course! Despite the challenging content, this course actually is taught at least at the same level as the ones by Andrew Ng, Daphne Koller, and Geoffrey Hinton. Congratulations Martha and Adam! You are awesome and are my heroes! Thanks a lot! George
The material is very good. But this course needs better instructors/ method of teaching. The book is also written in an unnecessarily technical way filled with jargon. explanations are not clear, simple stuff is presented in a very complicated manner for no obvious way.
Good content, but there was a highly unpleasant surprise in the programming assignments, namely this: "Retakes: You can attempt this assignment 5 times every 4 months." First of all, this is a highly unusual and therefore unexpected requirement on Coursera. Also, considering how buggy graders are and that some assignments require submitting results separately from the notebook, this is a really high risk of having to wait 4 months for another chance.
Needs more work in my opinion. It's not bad of course. I just believe that more intuition should be built with better examples, outside the text book rather than going through the actual mathematical proofs
Great course from an instructor/pedagogy perspective; HORRIBLE support from the coursera team. I have been prevented from completing the course for over 3 weeks now due to an issue that they will not help me with.
Excellent.
BEFORE this course: I’ve done a number of Coursera courses before. Whilst they are good, the level of learning tends to be superficial.
THIS is the first time I've dona a series of course (a module). These are the best courses I’ve taken and (after 3) I now feel I have learnt a very significant amount. Below applies to all three courses.
I have seen someone criticize the course by saying ‘it is just them talking through the Sutton & Barto book’. In defense: (i) the book *does* seem to be *the* seminal introductory text, (ii) importantly, they have selected which bits to cut out, (iii) I have now read through the recommended chapters as part of the course and have far greater insight than if just reading the book myself.
In some cases, the slides show things clearer than in the book. In some cases, the sentences are far too complicated to digest oinne one go. You need to rewind again and again to understand things.
I have found the time taken to do the Python assignments to be much longer than they suggest but this is largely down to my lack of Python abilities. I lot of time was spent improving my Python – which was a good by-product. The intermediate checking of code (within Jupyter) could be better but the forums help.
The content is good. The lectures are poor with the instructors doing little explanation other than reading equations off slides.
Feels to be too focussed on theory and math, instead on practically applying the best techniques.
Excellent Course and Lectures. Loved it!! So important to read the chapters in the book ahead of time. Book is also excellent!! I liked the way the instructors explained the equations and broke them down. Nicely done!! I wish some more of the questions in the quiz reflected the data structures we use in the programming exercise, which will be super-helpful to reinforce the concepts when we do the programming exercises. In other words, an intermediate step of a worked example between the Pseudo-Code Algorithm in the Texbook/Lectures and the Programming Exercise. For example, more of the Feature_Vector -> Action_Value Calculation - even if we have to do some matrix manipulation by hand, that'd be wonderful. One of the quizzes has something like that (but more simplified) - which was perfect.
Very nice introduction to RL! The course follows tight the book from Sutton, and differently to other courses in Coursera, it actively requires reading it. This was a plus note! The quizzes were also inspiring and some of them required understanding beyond the lectures. Coding exercises were just OK to show the overall ideas. Do not expect to do a research project, it is still an online-almost-free course given by top academic fellows. Great experience. Thanks!
This was a good course but I really struggled to understand how each of the value functions translated into code.
Excellent instruction that guides through the core material of part 2 of Sutton & Barto's Reinforcement Learning: an Introduction. The instructors additionally teach complementary material not found in the book. The notebooks got me "making something real" with the material in a way that deepened my understanding beyond a theoretical/pen-and-paper treatment. I appreciate the care that went into setting up the RL learning environment, creating test cases, and visualizations of performance -- it's awesome when the agents come together and you can see how well they perform!
A couple very minor notes on the lectures. The pacing of speech by Dr. Adam White often felt stiff and clipped. In future video courses he might benefit from practicing changing his tone, speed, and pauses to sound more natural. Similarly, Dr. Martha White's microphone was positioned in such a way that her breathing between sentences is captured, and sounds pretty loud. Improving these aspects of presentation in the future can make the lectures flow more naturally and reduce some friction from the distractions.
Those are nits on an otherwise excellent course. Thank you very much for putting the materials together! See you in the next one!
The third course of the specialization is excellent and it provides a solid foundation on problems with arbitrarily large state spaces that rely on approximate solution methods. The lectures are very well explained. It’s strongly recommended to read each book chapter in advance before watching the lectures to be able to better understand the concepts and be able to answer the quizzes. The content in this course is quite abstract and it is heavily dependent on statistics and calculus. It was very nice to integrate reinforcement learning with neural networks as part of one of the assignments as well as to implement the swing-up pendulum. I am looking forward to begin the capstone project.
Excellent course. The videos, quizzes, and especially the exercises add a lot of extra value to the text book (which is available for free - Sutton and Burto, 2nd edition). Of course it is not perfect - the videos are sometimes a bit dry, the NN part was brushed over too quickly for a beginner (luckily I had taken some courses about deep learning, so I was ok - but if you don't know the basics of NN, week 2 might be quite challenging for you). Other than that the biggest disadvantage is that the course forums are still quite empty - and so if you get stuck you can be on your own... But you shouldn't get stuck, and I guess this will improve over time.
This, the third in an exceptionally well-paced series of four courses on Reinforcement Learning, extends the scope of the subject to include parameterized functions (i.e., neural networks). The section on tiling methods is especially interesting. The course is taught under the auspices of professors who, quite literally, wrote the book on reinforcement learning, and includes several video lectures by leading practitioners and theorists in the field. The final programming assignment, in particular, made me feel like I did when I wrote my first computer program that actually did what it was supposed to way back when -- delight and amazement.
Learned a lot through the course. This specialization is to teach you through the whole reinforcement learning textbook. Very informative, but the programming assignments are very difficult.
There are many tiny details to notice while programming, and the discussion forum is not very active. I suggest find some clues from the old posts when having difficulties, and ask / answer questions as much as possible to help yourself and the others.
Great course and specialization. The teachers are great, the material well presented and balanced. I strongly recommend this course to anyone interested in the field of Reinforcement Learning. For maximum chance of success I suggest following all 3 courses in succession and investing the necessary amount of time to read the textbook chapters as specified at the beginning of each week.
Looking forward to completing the capstone project now!
The course is very comprehensive on the content. But I think the difficulty of this course is in some sense too high for most people who don't have a background in engineering degree due to the extensive use of advanced mathematics. I think it might be a better idea if you are focusing on a few critical algorithms that trying to cover too much algorithms which is quite overwhelming
An almost overwhelming amount of material, however we managed to navigate through the thicket. The labs were well maintained and provided robust tests so that one could have a high degree of confidence in the solution before submitting to the grader. I really appreciate this. I would recommend this course to anybody wanting a serious introduction to reinforcement learning.
This course was really well put-together. Lectures were high quality in terms of both content and production. Learning objectives for the whole course, and for individual lectures were very clear. The labs were very well thought-out, interesting and fun, and were very relevant to the course material. Quizzes were also appropriate to the material, with good, clear questions.