Re: [Idnet] Architecture and standard opportunities of IDN

김민석 <> Tue, 01 August 2017 09:25 UTC

Return-Path: <>
Received: from localhost (localhost []) by (Postfix) with ESMTP id 3EBE1132CB7 for <>; Tue, 1 Aug 2017 02:25:37 -0700 (PDT)
X-Virus-Scanned: amavisd-new at
X-Spam-Flag: NO
X-Spam-Score: -1.9
X-Spam-Status: No, score=-1.9 tagged_above=-999 required=5 tests=[BAYES_00=-1.9, HTML_IMAGE_RATIO_06=0.001, HTML_MESSAGE=0.001, RP_MATCHES_RCVD=-0.001, SPF_PASS=-0.001] autolearn=ham autolearn_force=no
Received: from ([]) by localhost ( []) (amavisd-new, port 10024) with ESMTP id frP3qH55T846 for <>; Tue, 1 Aug 2017 02:25:35 -0700 (PDT)
Received: from ( []) by (Postfix) with ESMTP id 9416D132CB5 for <>; Tue, 1 Aug 2017 02:25:33 -0700 (PDT)
Received: from unknown (HELO ( by with ESMTP; 1 Aug 2017 18:25:36 +0900
Received: from ( by ( with Microsoft SMTP Server (TLS) id 14.3.319.2; Tue, 1 Aug 2017 18:25:29 +0900
Received: from ([]) by ([]) with mapi id 14.03.0319.002; Tue, 1 Aug 2017 18:25:27 +0900
From: =?utf-8?B?6rmA66+87ISd?= <>
To: yanshen <>, Pedro Martinez-Julia <>
CC: "" <>
Thread-Topic: Architecture and standard opportunities of IDN
Thread-Index: AdMGA6fVfWTKGPKySXypIgnjpRC5HgDmjdFxABe1v5AAKO8WiA==
Date: Tue, 1 Aug 2017 09:25:26 +0000
Message-ID: <>
References: <> <>, <>
In-Reply-To: <>
Accept-Language: ko-KR, en-US
Content-Language: ko-KR
X-MS-Has-Attach: yes
x-originating-ip: []
Content-Type: multipart/related; boundary="_004_5BC916BD50F92F45870ABA46212CB29C019C4954SMTP1etriinfo_"; type="multipart/alternative"
MIME-Version: 1.0
Archived-At: <>
Subject: Re: [Idnet] Architecture and standard opportunities of IDN
X-Mailman-Version: 2.1.22
Precedence: list
List-Id: "The IDNet \(Intelligence-Defined Network\) " <>
List-Unsubscribe: <>, <>
List-Archive: <>
List-Post: <>
List-Help: <>
List-Subscribe: <>, <>
X-List-Received-Date: Tue, 01 Aug 2017 09:25:37 -0000

Dear Yansen,

Thoes are good key points!

We firslty need to set real networking model to collect data (Qos Data), then making a scenario which component of data-set can be applied with ML algorithms such as RL, RNN and etc.

It's a little hard to consider key characters for ML-based data structure to be clear, but I can make couple of opinions to the key characters of structure in aspect of ML algorithm.

1) QoS may be convert to reward for RL, but it needs learning processing time. We can set an initial period to collect information for learning process. In general, an initial period should be defined by random process in RL. After then, the data will be continuously updated.

2) Set a policy. Policy is important points for RL. It has all of the next available values for actions so that the policy is a brain involved with opimized data.

3) We should make clear of purpose for "Source-destination". Sheng has a plan to organize a dedicated session in next NMRG meeting so it may be clear for intelligent management before establishment of ML-based data structures.


Min-Suk Kim

Senior Researcher / Ph.D.

보낸 사람 : "yanshen" <>
보낸 날짜 : 2017-07-31 22:15:37 ( +09:00 )
받는 사람 : 김민석 <>kr>, Pedro Martinez-Julia <>
참조 : <>
제목 : RE: Architecture and standard opportunities of IDN

Dear Kim,

May you have seen that we have a consensus on the potential standardization point of data and data format.

In your experience, what the structure or feature that ML-based algorithm potentially tend to? If we have a consensus and full view in this question, the data structure may be clear.

For example, if we want to train a model with QoS data (or reward). Assume that these data can be captured via some method. Some following key characters of data structure may be needed:

-          QoS data, such as delay. It may be organized with time series/quantized (the “reward”) / ...

-          Starting-ending time. It is significant and should be differentiated.

-          Time interval. We can capture the delay values in different frequency. High frequency time series delay value may be “down-sampled” when it is used in training low frequency model.

-          Source-destination. Obviously.

-          TBD

Any other keys?


From: 김민석 []

Sent: Monday, July 31, 2017 10:06 AM

To: Sheng Jiang <>om>;

Cc: yanshen <>

Subject: RE: Architecture and standard opportunities of IDN

Dear Sheng,

Thank  you for giving us last IDNET side-meeting in Prague.

Many guys mentioned data-set problems in the IDN archetecture, so I strongly agree with these kinds of opinions.

BUT, we should carefully consider a ML-based algorithm set-up for Reinforcement Learning (RL) that you described in the other slide. Reward is one of the components in RL algorithm so that we should take all of the options of RL into consideration, such as which action we can take in the loop, which state should be optimized for reward, how policy will cover and dominate for learning process.

I submitted and presented our related internet draft based on RL in NMRG, Prague so that it might help to figure out more specifically. Hopefully. we are additionally able to discuss related RL setting-up more by email discussion or next meeting.


Min-Suk Kim

Senior Researcher / Ph.D.

보낸 사람 : "Sheng Jiang" <<>>
보낸 날짜 : 2017-07-26 20:38:50 ( +0900 )
받는 사람 :<> <<>>
참조 : yanshen <<>>
제목 : [Idnet] Architecture and standard opportunities of IDN

Hi, all,

Please find the below figure that I shared in the IDN discussion meeting last week in Prague. This illustrates the IDN architecture in a very generic way. The on-site participants had the very similar understanding that the upper left box – the unified and selected data is should be the prior work towards standardization. Also, the participants shared the same opinion that the data structure may be various among use cases. So, we should keep discussing on use cases with the target to converge on one or two significant and specific use cases.