[mlpack] GSOC 18 [Mlpack] : Reinforcement learning

[Marcus Edel]

Hello Rohan,

thanks for the update.

> Presently, I am going through the codes of mlpack. I had a basic idea of
> template meta-programming , hence I am comfortable with the codes. I will try to
> implement the Policy gradient method soon. However I request you to mention if
> there are any issues which can be resolved now.

Sounds good, we do our best to add more issues over the next weeks, that are
project related and are helpful to get familiar with the codebase.

Thanks,
Marcus

> On 17. Feb 2018, at 08:05, Rohan Raj <rajrohan1108 at gmail.com> wrote:
> 
> Hello Marcus,
> 
> I sincerely appreciate your quick respnse.
> 
> I would like to mention that I have used CNN in various of my projects. These can easily be found in my github <https://github.com/luffy1996> page. At UIUC, I used my network on top of pretrained alex network. Hence, I have had the flavour of working with CNN before. I sincerely believe that CNN will have positive impact in the problem being discussed.  This has already been shown in various atari games which were trained from pixels.
> 
> I do understand that the working basics of different packages differ significantly . I have experienced this before as I have worked on tensorflow, theano and caffe in past. However what I find is that, the logic behind the implementation is essentially the same. All machine learning packages have been designed to implement the theory of deep learning/machine learning , which is essentially the same. Hence I believe that mlpack is not much different from theano theoretically.
> 
> Presently, I am going through the codes of mlpack. I had a basic idea of template meta-programming , hence I am comfortable with the codes. I will try to implement the Policy gradient method soon. However I request you to mention if there are any issues which can be resolved now.
> 
> I would also like to mention that I go by the name of luffy1996  <https://github.com/luffy1996>in the irc channel.
> ᐧ
> 
> Rohan Raj
> Department of Chemical Engineering
> Indian Institute of Technology Guwahati
> Assam , India
> Phone : +91 8723990557, +91 8651776581
> 
> ᐧ
> 
> Rohan Raj
> Department of Chemical Engineering
> Indian Institute of Technology Guwahati
> Assam , India
> Phone : +91 8723990557, +91 8651776581
> 
> 
> 
> On 17 February 2018 at 11:59, Rohan Raj <rajrohan1108 at gmail.com <mailto:rajrohan1108 at gmail.com>> wrote:
> Hello Marcus,
> 
> I sincerely appreciate your quick respnse.
> 
> I would like to mention that I have used CNN in various of my projects. These can easily be found in my github <https://github.com/luffy1996> page. At UIUC, I used my network on top of pretrained alex network. Hence, I have had the flavour of working with CNN before. I sincerely believe that CNN will have positive impact in the problem being discussed.  This has already been shown in various atari games which were trained from pixels.
> 
> I do understand that the working basics of different packages differ significantly . I have experienced this before as I have worked on tensorflow, theano and caffe in past. However what I find is that, the logic behind the implementation is essentially the same. All machine learning packages have been designed to implement the theory of deep learning/machine learning , which is essentially the same. Hence I believe that mlpack is not much different from theano theoretically.
> 
> Presently, I am going through the codes of mlpack. I had a basic idea of template meta-programming , hence I am comfortable with the codes. I will try to implement the Policy gradient method soon. However I request you to mention if there are any issues which can be resolved now.
> 
> I would also like to mention that I go by the name of luffy1996  <https://github.com/luffy1996>in the irc channel.
> ᐧ
> 
> Rohan Raj
> Department of Chemical Engineering
> Indian Institute of Technology Guwahati
> Assam , India
> Phone : +91 8723990557, +91 8651776581
> 
> 
> 
> On 16 February 2018 at 17:18, Marcus Edel <marcus.edel at fu-berlin.de <mailto:marcus.edel at fu-berlin.de>> wrote:
> Hello Rohan,
> 
> thanks for getting in touch.
> 
>> I have a good knowledge in neural networks and deep learning.Previous summer, I
>> did my summer internship at Beckmann Institute, UIUC (University of Illinois at
>> Urbana-Champaign), USA on deep learning in cancer imaging.
> 
> That sounds really interesting, would be awesome if Deep learning would have an
> positive impact on this important problem, I think you used some CNN flavor in
> your experiments?
> 
>> Previous semester , I took Computer Vision using machine learning course at my
>> college. I proposed a transfer learning architecture for semantic segmentation
>> in deep learning as a semester project. The codes can be found here.
> 
> This looks really interesting as well, note Theano is somewhat different from
> what we usally do at mlpack.
> 
>> Presently I am going through the code structure of mlpack. I am comfortable with
>> the software because I have good background in C++. Since there are none tickets
>> open presently, I am currently following Marcus's suggestion to go through the
>> code base and try to improve the codes. I will be grateful to any member who
>> would like to provide any suggestions.
> 
> Another idea is to implement a simple RL method like (stochastic) Policy
> Gradients and test it on the existing environments, but don't feel obligated.
> 
> Let me know if I should clarify anything.
> 
> Thanks,
> Marcus
> 
> 
>> On 16. Feb 2018, at 07:29, Rohan Raj <rajrohan1108 at gmail.com <mailto:rajrohan1108 at gmail.com>> wrote:
>> 
>> Hello Everyone,
>> 
>> I am Rohan Raj , a pre-final year undergraduate student from IIT Guwahati.
>> 
>> I am doing my undergraduate research in artificial intelligence , focusing in deep reinforcement learning. Recently, I have submitted my research work, 'Weighted Experience Replay for Independent Q Learning in Multi-Agent Reinforcement Learning' , in ICML 2018 . 
>> 
>> I have a good knowledge in neural networks and deep learning.Previous summer, I did my summer internship at Beckmann Institute, UIUC (University of Illinois at Urbana-Champaign), USA on deep learning in cancer imaging.
>> 
>> Previous semester , I took Computer Vision using machine learning course at my college. I proposed a transfer learning architecture for semantic segmentation in deep learning as a semester project. The codes can be found here <https://github.com/luffy1996/transfer-learning-semantic-segmentation>.   
>> 
>> My blogs are regularly followed by various researchers in the world. You may like to read my introductory blogs on LSTMs <https://rohanrajblogs.blogspot.in/2016/12/writing-simple-lstm-model-on-keras.html> and supercomputer param isham <https://rohanrajblogs.blogspot.in/2017/01/supercomputer-param-ishan.html>.
>> 
>> I have been through the idea list and I am interested in working in reinforcement learning module. I have sufficient knowledge of DDQN networks and actor-critic networks. I have fairly good understanding of the PPO algorithms.
>> 
>> Presently I am going through the code structure of mlpack. I am comfortable with the software because I have good background in C++. Since there are none tickets open presently, I am currently following Marcus's suggestion to go through the code base and try to improve the codes. I will be grateful to any member who would like to provide any suggestions. 
>> 
>> You may want to have a look at my resume, which is attached with this email.
>> 
>> Thank You,
>> Rohan Raj
>> Indian Institute of Technology Guwahati
>> Assam , India
>> Phone : +91 8723990557.
>> 
>> 
>> 
>> ᐧ
>> <rohanraj_IIT_Guwahati_.pdf>
> 
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/20180218/820a0721/attachment-0001.html>