[Robot-learning] concepts in deep learning networks

[Nakul Gopalan]

The work seems to be a combination of AMDPs and the DAQN work that Mel and
Chris are working on. An end to end task is broken into subtasks, each with
an independent reward function and state-action space. They are learning
each subtask independently and not all at the same time like DAQN. They
chose a domain that Deepmind did for stacking blocks and showed that a
policy for stack blocks can be learned 45x faster. Their video page does
not exist so I can't really see their robot in action.
The paper itself lacks a lot of related work, they only cite the early
options work, ignore everything after that and criticise two papers from
last year. They do not point out that the hierarchy looks like a MAXQ
hierarchy and do not talk about optimality issues of such an approach.
Best
nakul


On 26 September 2017 at 15:00, Littman, Michael <mlittman at cs.brown.edu>
wrote:

> indeed... has anyone looked over the paper? any insights?
>
> On Tue, Sep 26, 2017 at 2:48 PM, Marie desJardins <mariedj at umbc.edu>
> wrote:
>
>> Looks an awful lot like AMDPs...
>>
>> https://www.forbes.com/sites/aarontilley/2017/09/19/ai-start
>> up-invents-trick-for-robots-to-more-efficiently-teach-themse
>> lves-complex-tasks/#5a6254d315fe
>>
>> https://drive.google.com/file/d/0B7-VbSZ5FzXBdURHLXV4OU9EOTQ/view
>>
>> Marie
>>
>> --
>> Dr. Marie desJardins
>> Associate Dean for Academic Affairs
>> College of Engineering and Information Technology
>> University of Maryland, Baltimore County
>> 1000 Hilltop Circle
>> Baltimore MD 21250
>>
>> Email: mariedj at umbc.edu
>> Voice: 410-455-3967 <(410)%20455-3967>
>> Fax: 410-455-3559 <(410)%20455-3559>
>>
>> _______________________________________________
>> Robot-learning mailing list
>> Robot-learning at cs.umbc.edu
>> https://lists.cs.umbc.edu/mailman/listinfo/robot-learning
>>
>>
>
> _______________________________________________
> Robot-learning mailing list
> Robot-learning at cs.umbc.edu
> https://lists.cs.umbc.edu/mailman/listinfo/robot-learning
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cs.umbc.edu/pipermail/robot-learning/attachments/20170927/2e08f81e/attachment.html>