[mlpack] GSoC 2016 Project: Neuroevolution algorithms
Marcus Edel
marcus.edel at fu-berlin.de
Thu Mar 17 10:47:45 EDT 2016
Hello Nakul,
> Apologies for getting back so late, I was working on the draft for the proposal.
> I have shared the proposal with mlpack and would appreciate your inputs.
Thanks for the proposal, we will take a look on the draft in the next days and
make comments.
> At the time Udit implemented the perceptron there wasn't any code to implement
> the perceptron using the FNN class.
>
> If there is a need to rewrite the perceptron using the fnn interface, I would
> work on that too.
If you like you can rewrite the perceptron using the ann modules.
> After your mention of the pole balancing problem, I read more about it and it is
> regarded as a pseudo-standard benchmark test for neuroevolution algorithms, so I
> would implement that as part of testing of the algorithm.
Sounds great, there are a couple of other tests that could be interessting like:
the mountain car task.
Thanks,
Marcus
> On 17 Mar 2016, at 14:11, Nakul Gulati <nakgulati at gmail.com> wrote:
>
> Hey Marcus,
>
> Apologies for getting back so late, I was working on the draft for the proposal. I have shared the proposal with mlpack and would appreciate your inputs.
>
> At the time Udit implemented the perceptron there wasn't any code to implement
> the perceptron using the FNN class.
>
> If there is a need to rewrite the perceptron using the fnn interface, I would work on that too.
>
> Or the other way. Take a look at the pole balancing problem that could be a neat
> test case.
>
> After your mention of the pole balancing problem, I read more about it and it is regarded as a pseudo-standard benchmark test for neuroevolution algorithms, so I would implement that as part of testing of the algorithm.
>
> Regards,
> Nakul
>
> On Sun, Mar 13, 2016 at 8:25 PM, Marcus Edel <marcus.edel at fu-berlin.de <mailto:marcus.edel at fu-berlin.de>> wrote:
> Hello Nakul,
>
>> I have gone through the literature and have a better understanding of the
>> algorithms and also have been spending time with the source code. It took a
>> little time but now I am acquainted with the source code. After understanding
>> the ann code I went ahead to check out the perceptron code to see it in action
>> but to my surprise perceptron although being a feed forward single layer neural
>> network doesn't use fnn code, is there a particular reason for it, am I missing
>> something?
>
> At the time Udit implemented the perceptron there wasn't any code to implement
> the perceptron using the FNN class.
>
>> Also about the GSoC project I feel confident with CNE and NEAT but like you had
>> mentioned the importance of test cases I would the priority would be to
>> implement CNE with test cases and then move to NEAT.
>
> Or the other way. Take a look at the pole balancing problem that could be a neat
> test case.
>
> I hope this is helpful, let me know if I can clarify anything.
>
> Thanks,
> Marcus
>
>
>> On 13 Mar 2016, at 11:53, Nakul Gulati <nakgulati at gmail.com <mailto:nakgulati at gmail.com>> wrote:
>>
>> Hey Marcus,
>>
>> I have gone through the literature and have a better understanding of the algorithms and also have been spending time with the source code. It took a little time but now I am acquainted with the source code. After understanding the ann code I went ahead to check out the perceptron code to see it in action but to my surprise perceptron although being a feed forward single layer neural network doesn't use fnn code, is there a particular reason for it, am I missing something?
>>
>> Also about the GSoC project I feel confident with CNE and NEAT but like you had mentioned the importance of test cases I would the priority would be to implement CNE with test cases and then move to NEAT.
>>
>> Regards,
>> Nakul Gulati
>> Website <https://nakulgulati.com/> || LinkedIn <http://in.linkedin.com/in/nakulgulati> || GitHub <https://github.com/nakulgulati/>
>>
>> On Thu, Mar 10, 2016 at 7:17 PM, Marcus Edel <marcus.edel at fu-berlin.de <mailto:marcus.edel at fu-berlin.de>> wrote:
>> Hello Nakul,
>>
>> It is a good decision to start with something easy that could be CNE or NEAT.
>> Afterwards, we can use that as a baseline for comparison with other
>> implementation. You should keep in mind, that you have to write tests, for every
>> evolution algorithm you write, and that often takes more time than the actual
>> implementation. Anyway, this year google made draft proposals part of the
>> proposal workflow. So we can comment on your timeline etc. and give feedback
>> once you submitted your application. Also take a look at:
>> http://write.flossmanuals.net/gsocstudentguide/ <http://write.flossmanuals.net/gsocstudentguide/>
>>
>> I hope this is helpful. Let me know if I can clarify anything,
>>
>> Thanks,
>> Marcus
>>
>>> On 10 Mar 2016, at 04:12, Nakul Gulati <nakgulati at gmail.com <mailto:nakgulati at gmail.com>> wrote:
>>>
>>> Hi,
>>>
>>> I am a final year Computer Science and Engineering student. I am interested in contributing to the project Neuroevolution algorithms as part of GSoC 2016. Some of the courses completed which are relevant to this project are: Data Structures and Design and Analysis of Algorithms, Soft Computing and Artificial Neural Networks.
>>>
>>> In order to get better understanding about the project and mlpack, the following steps were taken:
>>> Compiled and tested mlpack on OS X
>>> Compiled and tested nes emulator communication code: during this I ended up crashing the emulator hosted on mario.urgs.org <http://mario.urgs.org/> 4561, link to issue <https://github.com/zoq/nes/issues/1>. Which was promptly fixed by Marcus Edel, big shout out to him.
>>> Currently I am studying the reading material, starting with the CNE algorithm.
>>>
>>> For scope of GSoC project I propose to start with the implementation of the CNE algorithm at first and then move to a second (name would be specified soon) if time permits.
>>>
>>> I have little trouble setting up an exact timeline for the project. I would love to hear your insight about the approach and timeline and if you think what I propose is realistic.
>>>
>>> --
>>> Regards,
>>> Nakul Gulati
>>> Website <https://nakulgulati.com/> || LinkedIn <http://in.linkedin.com/in/nakulgulati> || GitHub <https://github.com/nakulgulati/>_______________________________________________
>>> mlpack mailing list
>>> mlpack at cc.gatech.edu <mailto:mlpack at cc.gatech.edu>
>>> https://mailman.cc.gatech.edu/mailman/listinfo/mlpack <https://mailman.cc.gatech.edu/mailman/listinfo/mlpack>
>>
>>
>>
>>
>> --
>> Regards,
>> Nakul Gulati
>> Website <https://nakulgulati.com/> || LinkedIn <http://in.linkedin.com/in/nakulgulati> || GitHub <https://github.com/nakulgulati/>
>
>
>
> --
> Regards,
> Nakul Gulati
> Website <https://nakulgulati.com/> || LinkedIn <http://in.linkedin.com/in/nakulgulati> || GitHub <https://github.com/nakulgulati/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20160317/1c5c15b1/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 5233 bytes
Desc: not available
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20160317/1c5c15b1/attachment-0001.bin>
More information about the mlpack
mailing list