Re: [Idnet] Benefits of Introducing AI into Network

On 11/05/2017 21:06, David Meyer wrote:
> Brian,
> 
> 
> On Wed, May 10, 2017 at 9:35 PM, Brian E Carpenter <
> brian.e.carpenter@gmail.com> wrote:
> 
>> On 10/05/2017 18:26, Fang, Luyuan wrote:
>>> I think what is driving all these points is scale. Without intelligence,
>> very large scale networks simply become unmanageable,
> 
> 
>> I am not sure that statement is always true. It is safe to say that
>> without *distributed solutions*, very large scale networks become
>> unmanageable, and AI technologies can be used to support distributed
>> solutions.
>>
> 
> Can you support this assertion? Further, what is an "AI technology",

I'm not brave enough to try to answer that in a few words. But clearly
ML is an example. Some kind of inference engine is another. Also, specific
heuristic algorithms seem very suitable for use in systems of distributed
agents, because each agent needs to respond to local conditions.

> and
> which ways can "AI technologies can be used to support distributed
> solutions."?

Either by installing heuristics in individual agents, or by allowing
agents to consult a central oracle when they need help ("AI as a service"
if you like). But there's a lot to work out.

>>
>> Let's take a specific example: the CASM work in the IETF. One approach to
>> massive scale support of address and prefix management is a massive
>> centralised database and an old-fashioned (NETCONF/YANG style) mechanism to
>> push allocations out into the network. Another approach is a large set of
>> distributed autonomic prefix managers that request new space from a
>> distributed address pool when necessary. The second one is definitely a
>> target for machine learning, in order to set optimal parameters for the
>> individual prefix managers.
>>
> 
> Phrases like "massive scale" don't seem to convey anything quantitative (or
> even qualitative, for that matter) about the problem space, and as such
> don't appear to be that useful (unless of course you can meaningfully
> quantify what is meant by "massive scale"). 

Well, in this case think of an ISP with tens of millions of subscribers,
and work back from that to how many distributed agents might be involved.
For example, at least 1000 agents each handling address space delegation
to 10000 subscribers.

> In addition, what kinds of
> machine learning would be applicable to the case of distributed autonomic
> prefix managers that you describe above (simply stating that it is so isn't
> that helpful)? For example, how would learning work, e.g., what kinds of
> models and related hyper-parameters do you envision being appropriate, and
> what kinds of data sets would be needed to efficiently train these models?
> In addition, how would inference work, where inference here means whatever
> you do with the trained model, including inference (computing the
> posterior), prediction (MAP or MLE), regression, ...?

Dunno. These are good questions.

>>
>>> and cannot maintain service availability. Just adding work force would
>> not help.
>>
>> It will help where the problem can genuinely be tackled by "divide and
>> conquer", for example where localised optimisation and localised repair can
>> help. But when localised solutions are inadequate and traditional
>> centralised solutions do not scale, AI can be the answer.
>>
> 
> What is AI in this context? In addition, why can AI (whatever that is) be
> the answer to localized or centralized solutions that don't scale?  Again,
> simply stating that it is so doesn't really help.

Learning from experience is the key benefit, I think. That's why I imagine
that the distributed agents could share a central ML oracle, which will
build up knowledge over the long term. To invent a heuristic out of thin
air: "An agent will experience a burst of address space requests at 19:00
in its local time zone. Therefore it should request extra space at 18:30."
That seems like something that an ML/inference engine could produce from
reports sent in by distributed agents.

Regards
   Brian

> Thanks,
> 
> Dave
> 
> 
>>> I would suggest to add
>>>
>>> -          ability to manage network at scale (as 2nd point)
>>>
>>> -          potential self-healing (after “predictive”)
>>
>> Agreed.
>>
>>     Brian
>>>
>>> Thx,
>>> Luyuan
>>>
>>> From: IDNET <idnet-bounces@ietf.org> on behalf of Sheng Jiang <
>> jiangsheng@huawei.com>
>>> Date: Tuesday, May 9, 2017 at 7:53 PM
>>> To: "idnet@ietf.org" <idnet@ietf.org>
>>> Subject: [Idnet] Benefits of Introducing AI into Network
>>>
>>> In ETSI NGP ISP, we have an Work Item for IDN (Intelligence-Defined
>> Network). In one of the general sections, We briefly describes the benefits
>> of introducing AI into network, as below.
>>>
>>>
>>> -          Towards Fully Autonomic Network
>>>
>>> -          Ability of Handling Complex Issues
>>>
>>> -          More Adaptive and Flex
>>>
>>> -          Predictive
>>>
>>> -          Potential Self-Evolving Ability
>>>
>>> -          Potential Decision Efficiency
>>>
>>> If there are major benefits that we have not covered, we would like to
>> include.
>>>
>>> Regards,
>>>
>>> Sheng
>>>
>>>
>>>
>>> _______________________________________________
>>> IDNET mailing list
>>> IDNET@ietf.org
>>> https://www.ietf.org/mailman/listinfo/idnet
>>>
>>
>> _______________________________________________
>> IDNET mailing list
>> IDNET@ietf.org
>> https://www.ietf.org/mailman/listinfo/idnet
>>
> 