[Dprglist] Method for humans to teach robots for deep learning

[Doug Paradis]

Ezra,
    Very interesting insights! I suspect it is used when the machine makes
an obvious error.

Regards,
Doug P.

On Fri, Feb 23, 2018 at 6:17 PM, Ezra Christensen <ezracc at gmail.com> wrote:

> Interesting.
>
> I wonder whether human's watching Google's latest DeepMind go through
> iterations to test and train itself in Go would have been able to know
> whether the bot was doing a good or bad job. It reportedly came up with new
> strategies that even expert players hadn't used, which allowed it to beat
> its predecessor. As that emerged, a human watching that may not have been
> able to recognize what the bot was doing and not rewarded such behavior.
>
> For human training, sometimes it's good to let a person do a bad job for a
> period of time so they learn how to differentiate between what good and bad
> is and why. For learning algo's, if you don't give it enough iterations it
> may not have a complete data set to figure out what to pay attention to and
> what to ignore, assuming it has access to relevant data.
>
> Does TAMER reach a point it starts ignoring the human feedback as it
> determines they don't really know what they're talking about. ;)
>
> Still cool.
>
>
>
> ------ Original Message ------
> From: "Doug Paradis" <paradug at gmail.com>
> To: "DPRG" <dprglist at lists.dprg.org>
> Sent: 2/23/2018 1:34:44 PM
> Subject: [Dprglist] Method for humans to teach robots for deep learning
>
> Interesting article:
> http://www.machinedesign.com/motion-control/good-robot-bad-
> robot-future-robotic-feedback-deep-learning?NL=MACD-001&
> Issue=MACD-001_20180223_MACD-001_551&sfvc4enews=42&cl=article_2_b&utm_rid=
> CPG05000003813138&utm_campaign=15533&utm_medium=email&elq2=
> ced390a9de014072adc478258295d454
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.dprg.org/pipermail/dprglist-dprg.org/attachments/20180223/507a41b2/attachment.html>