[mlpack] GSoC 17: Interested in the Reinforcement learning project

[Arun Reddy]

Hey Marcus,

On Wed, Mar 1, 2017 at 10:15 AM Marcus Edel <marcus.edel at fu-berlin.de>
wrote:

> Hello Arun,
>
> At first, Congratulations on being accepted for GSoC 2017.
>
>
> Thanks and welcome, looking forward to have a lot of fun over the summer.
>
Yes, Definitely it is going to be fun summer.


>
> I am Arun Reddy, thrid year PhD Student in machine learning at Arizona
> State
> University, USA. My current area of research is Transfer learning/Domain
> adaptation using Deep Learning, specifically on the problem "Is human
> expertise
> transferable?".
>
>
> That sounds really interesting, I guess, the "Reinforcement Learning"
> project
> idea goes kinda in a similar direction and would fit?
>
True, the reinforcement learning project is spot on. Given that the system
is trying to learn that expertise through exploration and reward system..
it will be interesting to see if that can be learnt and transferable to
other models. Long way.. but definitely aligns with my goals.


>
> I have a good understanding of Neural Networks & Reinforcement
> learning(RL), and
> would like to apply for the "Reinforcement Learning" project. I have done
> the
> relevant coursework at my university[1], and did the David Silver's
> course[2] as
> well. During the coursework, I learned how the agents interact with the
> environment and the underlying challenges through Edx Pacman projects[3]
> and
> also the implemented famous Atari Deep RL paper[4]. I am currently working
> in
> the direction of Reinforcement learning(RL) and adaptation, investigating
> if it
> is possible to improve the model learned by agents through interaction by
> scaffolding with existing models. Contributing to this project will help
> me to
> get a hands-on and a deep understanding of the existing DeepRL
> algorithms.  I am
> looking forward to contribute to mlpack, with a motive to get my hands
> dirty,
> learn to write efficient and maintainable code from scratch, and be part
> of the
> open source community.
>
>
> I didn't know about the Pac-Man project, the code examples and clear
> directions
> are really nice. Also, since you pointed out some really interesting
> references,
> have you seen "Deep Reinforcement Learning: An Overview" by Yuxi Li, it's a
> really comprehensive overview.
>
I never stumbled upon it. So exhaustive and well written. Going on to my
reading list. Thanks for sharing the reference.


>
> I was able to successfully compile the code and run few tests. Also got the
> gym_tcp_api working in my local environment. As suggested on the mailing
> list by
> Marcus, I would like to start off by contributing to few existing issues
> and
> move on to implementing policy gradients to get a hang of mlpack.
>
>
> Starting with a simple method like stochastic or deterministic Policy
> gradients
> is a really good idea, I think Temporal Difference Learning is another
> approach
> that might be manageable and interesting.
>
I have started with something on stochastic policy gradients.. i will look
into temporal difference learning too. Excited to be around.:)


>
> Thanks,
> Marcus
>
> On 28 Feb 2017, at 21:22, Arun Reddy <arunreddy.nelakurthi at gmail.com>
> wrote:
>
> Hello Devs and fellow GSoC enthusiasts,
>
> At first, Congratulations on being accepted for GSoC 2017.
>
> I am Arun Reddy, thrid year PhD Student in machine learning at Arizona
> State University, USA. My current area of research is Transfer
> learning/Domain adaptation using Deep Learning, specifically on the problem
> "Is human expertise transferable?".
>
> I have a good understanding of Neural Networks & Reinforcement
> learning(RL), and would like to apply for the "Reinforcement Learning"
> project. I have done the relevant coursework at my university[1], and did
> the David Silver's course[2] as well. During the coursework, I learned how
> the agents interact with the environment and the underlying challenges
> through Edx Pacman projects[3] and also the implemented famous Atari Deep
> RL paper[4]. I am currently working in the direction of Reinforcement
> learning(RL) and adaptation, investigating if it is possible to improve the
> model learned by agents through interaction by scaffolding with existing
> models. Contributing to this project will help me to get a hands-on and a
> deep understanding of the existing DeepRL algorithms.  I am looking forward
> to contribute to mlpack, with a motive to get my hands dirty, learn to
> write efficient and maintainable code from scratch, and be part of the open
> source community.
>
> I was able to successfully compile the code and run few tests. Also got
> the gym_tcp_api working in my local environment. As suggested on the
> mailing list by Marcus, I would like to start off by contributing to few
> existing issues and move on to implementing policy gradients to get a hang
> of mlpack.
>
> [1] http://rakaposhi.eas.asu.edu/cse571/
> [2] http://www0.cs.ucl.ac.uk/staff/d.silver/web/Teaching.html
> [3] http://ai.berkeley.edu/project_overview.html
> [4] https://arxiv.org/abs/1312.5602
>
>
> Happy coding,
> Arun
>
> _______________________________________________
> mlpack mailing list
> mlpack at lists.mlpack.org
> http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
>
>
>
Regards,
Arun
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/20170301/85490801/attachment-0001.html>