[mlpack] [GSoC] Asking About the Feasibility of an Idea in RL

[Marcus Edel]

Hello Ahmed,

sorry for the slow response, it's definitely a challenge to get everything
together in time, but if you are up for that I think this is a neat project to
work on. Let me know what you think.

Thanks,
Marcus

> On 8. Mar 2019, at 21:05, Ahmed Maher <iahmedmaher at gmail.com> wrote:
> 
> Sure! pointer-generator networks just have a layer on top of the normal pointer networks which makes the code Namrata96 is planning to finish before May (if I understood correctly) extendable to what I mentioned earlier; what would be left is an RL algorithm (ex: ACKTR which is already proposed or just simple REINFORCE which is implemented) and the MCTS. Is the scope of tackling both suitable for GSoC? Is that a lot or too little since you are more aware of how much time development for the mlpack library would need? Or do you think that it is not a good idea in the first place?
> 
> On Thu, Mar 7, 2019 at 11:07 PM Marcus Marcus <marcus at virtual-artz.de <mailto:marcus at virtual-artz.de>> wrote:
> Hello Ahmed,
> 
> there is a WIP PR about pointer-generator networks:
> https://github.com/mlpack/mlpack/issues/1725 <https://github.com/mlpack/mlpack/issues/1725>, which goes in a similair direction
> and here is the mail discussion:
> http://knife.lugatgt.org/pipermail/mlpack/2019-February/003888.html <http://knife.lugatgt.org/pipermail/mlpack/2019-February/003888.html>. Do you
> think we should combine both approaches?
> 
> Thanks,
> Marcus
> 
> 
>> On 7. Mar 2019, at 13:11, Ahmed Maher <iahmedmaher at gmail.com <mailto:iahmedmaher at gmail.com>> wrote:
>> 
> 
>> Dear all,
>> 
>>          I hope you all are having a good day. I am a computer engineering student at Cairo University, Egypt currently in my last year. 
>> 
>>          I want to ask the maintainers of mlpack about how interested are they in the idea of adding pointer networks to the code base. I find it a good idea because pointer networks trained using deep RL algorithms have proven to be really good at solving combinatorial optimization problems like TSP and even on par with tools like Google OR-Tools in VRP while being end-to-end! Since mlpack monopolizes on speed, I think it would be good to have a foundation for expanding in this direction adding maybe MCTS later on which will make mlpack a necessity for people wanting to solve combinatorial problems as speed could be a deal breaker for other tools.
>> 
>>          What are your thoughts on that? I attached the papers I have read about this topic for reference. Kindly find them below. Thanks.
>> 
>> Pointer Networks <https://arxiv.org/abs/1506.03134>
>> Neural Combinatorial Optimization with Reinforcement Learning <https://arxiv.org/abs/1611.09940>
>> Reinforcement Learning for Solving the Vehicle Routing Problem <https://arxiv.org/abs/1802.04240>
>> 
>> Best Regards,
>> -- 
>> Ahmed Maher AlAmir
>> Senior Student, Computer Engineering Department
>> Faculty of Engineering, Cairo University
>> Giza, Egypt
> 
>> _______________________________________________
>> mlpack mailing list
>> mlpack at lists.mlpack.org <mailto:mlpack at lists.mlpack.org>
>> http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack <http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack>
> 
> -- 
> Ahmed Maher AlAmir
> Senior Student, Computer Engineering Department
> Faculty of Engineering, Cairo University
> Giza, Egypt

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/20190314/58485085/attachment.html>