(The paper can be found here: http://conferences.sigcomm.org/hotnets/2014/papers/hotnets-XIII-final116.pdf )

Xiaoqi Yin is the presenter, he introduces himself more of a control theory person than a networking person.

In the current system, a bitrate controller resides in video client decides what video rate to request for given the network environment. There are many challenges for rate adaptation: it's hard to predict future bandwidth, interactions with TCP ... etc.

As a result -- there are 50+ papers in this field, but there is no way to compare one to another. Moving forward, the author argues that the community needs a systematic framework, instead of yet another point in the space. This work tries to provide such a framework, allowing us to systematically explore the design space and evaluating all these algorithms. Given the framework, they also further proposed an algorithm using control theory (MPC).

The proposed framework use a QoE model to evaluate the algorithms' QoE "score", and see (1) how sensitive are the algorithms to the operating environment?, and (2) How far the algorithms are from the optimal?

The authors classify existing algorithms into two classes: (1) rate-based (RB) algorithms and (2) buffer-based (BB) algorithms. They then compare the performance of RB, BB, and their proposed MPC algorithm.

In general, they found MPC > BB > RB. In other words, even though BB performs better than RB, MPC performs even better in term of QoE. However, both MPC and RB suffers if prediction error is large. Nevertheless, if the prediction error is bounded, then MPC can perform better than BB.

Other insights the talk provides including:

- All algorithm benefits from a finer-grained set of bitrates.

- MPC/BB can achieve near-zero buffering, compare to RB.

Q: (TY Huang from Netflix) - as an ABR algorithm designer, I really like the spirit of pursuing a framework for the algorithm design. However, the solution you proposed is heavily depends on the QoE model, which is still not well-understood yet. Any comment on how we can go forward without a well-understood QoE model with this framework?

A: The QoE model we proposed is a pretty general linear combination of some very well-known factors. The designers can play with the weights between factors to suit for different personal preference.

Q: (an audience from UIUC) Since video has a rate associated with them, should we use UDP now?

A: As I said in the beginning, I am not a networking person (laughter in the room), but the framework should be general enough to work with different transport algorithms.

Q: (Bruce Magg) Here you talk about VoD, how about live streaming? For live streaming, jitter matters too. How do we know what should be counted into the objective function?

A: It is indeed very important to incorporate domain knowledge into the objective function.

Q: (Dave Oran from Cisco) There is no commercial player that doesn't use buffer occupancy, it is always a hybrid model. Also prediction error is also correlated to the number of flows share the bottleneck link. How can you incorporate those information into your framework?

A: We can further improve the model if there is more information available.

Q: (Nina Taft from Google) What's the function you use as the objective function? They seems to be the key to your framework.

A: The QoE function we use is a linear combination between several key factors, such as rebuffer rates and video rates.

Q: (TY Huang from Netflix) I have a question regards to the usage of linear combination. How do you deal with confounding factors in this case?

A: We take care of confounding factors by taking both buffer occupancy and video rate into account.

Xiaoqi Yin is the presenter, he introduces himself more of a control theory person than a networking person.

The question this paper tries to address is: how to design a dynamic bitrate adaptation algorithm that provides better QoE, and how to do it in a systematic way.

In the current system, a bitrate controller resides in video client decides what video rate to request for given the network environment. There are many challenges for rate adaptation: it's hard to predict future bandwidth, interactions with TCP ... etc.

As a result -- there are 50+ papers in this field, but there is no way to compare one to another. Moving forward, the author argues that the community needs a systematic framework, instead of yet another point in the space. This work tries to provide such a framework, allowing us to systematically explore the design space and evaluating all these algorithms. Given the framework, they also further proposed an algorithm using control theory (MPC).

The proposed framework use a QoE model to evaluate the algorithms' QoE "score", and see (1) how sensitive are the algorithms to the operating environment?, and (2) How far the algorithms are from the optimal?

The authors classify existing algorithms into two classes: (1) rate-based (RB) algorithms and (2) buffer-based (BB) algorithms. They then compare the performance of RB, BB, and their proposed MPC algorithm.

In general, they found MPC > BB > RB. In other words, even though BB performs better than RB, MPC performs even better in term of QoE. However, both MPC and RB suffers if prediction error is large. Nevertheless, if the prediction error is bounded, then MPC can perform better than BB.

Other insights the talk provides including:

- All algorithm benefits from a finer-grained set of bitrates.

- MPC/BB can achieve near-zero buffering, compare to RB.

**Q&A:**Q: (TY Huang from Netflix) - as an ABR algorithm designer, I really like the spirit of pursuing a framework for the algorithm design. However, the solution you proposed is heavily depends on the QoE model, which is still not well-understood yet. Any comment on how we can go forward without a well-understood QoE model with this framework?

A: The QoE model we proposed is a pretty general linear combination of some very well-known factors. The designers can play with the weights between factors to suit for different personal preference.

Q: (an audience from UIUC) Since video has a rate associated with them, should we use UDP now?

A: As I said in the beginning, I am not a networking person (laughter in the room), but the framework should be general enough to work with different transport algorithms.

Q: (Bruce Magg) Here you talk about VoD, how about live streaming? For live streaming, jitter matters too. How do we know what should be counted into the objective function?

A: It is indeed very important to incorporate domain knowledge into the objective function.

Q: (Dave Oran from Cisco) There is no commercial player that doesn't use buffer occupancy, it is always a hybrid model. Also prediction error is also correlated to the number of flows share the bottleneck link. How can you incorporate those information into your framework?

A: We can further improve the model if there is more information available.

Q: (Nina Taft from Google) What's the function you use as the objective function? They seems to be the key to your framework.

A: The QoE function we use is a linear combination between several key factors, such as rebuffer rates and video rates.

Q: (TY Huang from Netflix) I have a question regards to the usage of linear combination. How do you deal with confounding factors in this case?

A: We take care of confounding factors by taking both buffer occupancy and video rate into account.