Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
  • Sep 18 11:24
    basnijholt edited #293
  • Sep 17 21:58
    basnijholt edited #293
  • Sep 17 21:57
    basnijholt edited #293
  • Sep 17 21:56
    basnijholt synchronize #293
  • Sep 17 21:56

    basnijholt on pre-commit-bumps

    bump pre-commit filter dependen… (compare)

  • Sep 17 21:55
    basnijholt opened #293
  • Sep 17 21:54

    basnijholt on pre-commit-bumps

    bump pre-commit filter dependen… (compare)

  • Sep 17 21:51

    basnijholt on master

    remove scikit-learn dep and pin… (compare)

  • Sep 17 21:37

    basnijholt on fix-docs

    (compare)

  • Sep 17 21:36

    basnijholt on master

    release strict pinning add BokehJS and fix jupyter_sph… remove _inline_js=False in adap… and 5 more (compare)

  • Sep 17 21:36
    basnijholt closed #291
  • Sep 17 21:35
    basnijholt synchronize #291
  • Sep 17 21:35

    basnijholt on fix-docs

    leave loky as deafult, change b… (compare)

  • Sep 17 21:33
    basnijholt edited #291
  • Sep 17 21:13
    basnijholt synchronize #291
  • Sep 17 21:13

    basnijholt on fix-docs

    fix syntax error (compare)

  • Sep 17 21:10
    basnijholt synchronize #291
  • Sep 17 21:10

    basnijholt on fix-docs

    fix syntax error (compare)

  • Sep 17 21:06
    basnijholt synchronize #291
  • Sep 17 21:06

    basnijholt on fix-docs

    fix syntax error (compare)

kwant-bot
@kwant-bot
alvarogi:
There is still some work to do, as the learner cannot run in parallel yet. I was planning to implement that during next week (I think I should have enough time), and then we can PR.
alvarogi:
About the report, I was thinking about combining the main notebooks as you said
alvarogi:
I only need to hand in the "first page of the report", so that should be fine
anton:
Yep.
kwant-bot
@kwant-bot
alvarogi:
Latest updates regarding this: summary and full discussion.
basnijholt:
Are there many hyper parameters that need to be tuned?
anton:
One (δ)
alvarogi:
There are actually 6, but 5 of them work with the default values. So in practice only one
anton:
δ should also work with the default values, wouldn't it?
alvarogi:
Indeed, but it is the parameter that has the strongest effect on the behavior of the learner
kwant-bot
@kwant-bot
alvarogi:
Let's say that someone wants to drastically reduce the uncertainty of the existing points rather than exploring new ones. Then it is enough to tune δ and leave the rest of parameters untouched.
kwant-bot
@kwant-bot
alvarogi:
Some news about the project:
  1. The final version of the AverageLearner1D is ready, in principle. I made some changes and it can now be run in parallel using the same commands as the Learner1D (see the tutorial).
  2. I will soon prepare a pull request to conclude the project (I will be quite busy working on a PhD proposal during the next two weeks, so I will do it after June 15 if that's ok).
  3. Should we present this work in a group meeting?
kwant-bot
@kwant-bot
anton:
Let's plan the presentation outside of the group meeting.
kwant-bot
@kwant-bot
alvarogi:
Okay, still only for the group? I am free any day after handing in the proposal on June 15.
kwant-bot
@kwant-bot
alvarogi:
I am all set with the proposal! I will work during this week on the pull request and the presentation to conclude the project. For which public should I aim the presentation?
anton:
Minimally our group, but you are free and welcome to announce it broader (e.g. the theory group, qutech, etc)
alvarogi:
Sounds good. Any preferences for date/time?
anton:
Yes.
kwant-bot
@kwant-bot
basnijholt:
I am excited for the PR :)
kwant-bot
@kwant-bot
alvarogi:
It is done!
kwant-bot
@kwant-bot
kwant-bot
@kwant-bot
alvarogi:
What about Friday at 17h?
anton:
I think it's best to schedule so that we'd end before 17. Working days and all...
alvarogi:
16h then?
anton:
Sounds good to me. @michaelwimmer does that work for you as well? (for the assessment)
kwant-bot
@kwant-bot
michaelwimmer:
This Friday?
anton:
yes
michaelwimmer:
So like 16-17?
alvarogi:
Probably a bit shorter, but yes
michaelwimmer:
Then it's fine\
anton:
Great, I'll add it to the group's calendar.
anton:
CC also @basnijholt and @jbweston if you want to join the presentation
kwant-bot
@kwant-bot
alvarogi:
Is it okay if I use Figure 1 from the paper for tomorrow's presentation? I will give a brief introduction to adaptive and that figure is quite useful
kwant-bot
@kwant-bot
anton:
Sure thing
kwant-bot
@kwant-bot
basnijholt:
Will the presentation be recorded? I can't make it unfortunately.
kwant-bot
@kwant-bot
anton:
Sure
kwant-bot
@kwant-bot
basnijholt:
Great presentation @alvarogi. I now understand better what your code does :)
basnijholt:
@anton, so the uncertainty of the average doesn't appear in the loss after all? Or am I missing something?
basnijholt:
The implementation seems quite simple in the end BTW. Although it was probably a lot of work to arrive at the right strategy.
basnijholt:
I wonder, you calculate the error using the student t-distribution, is it very different from the standard error of the mean instead (which is what I tried)?
basnijholt:
My implementation didn't work well in the end because it was too hard to tune the hyper parameters IRL.
kwant-bot
@kwant-bot
basnijholt:
I wonder how sensitive everything is to this delta. And do you have any intuition on how I should go about setting it?
basnijholt:
My application is one where one function realization takes 1 hour and I need ~2000 points to get a good average. So experimenting with parameters is expensive :P
kwant-bot
@kwant-bot
anton:
Good point. I think we can rewrite it as a part of loss.
anton:
Without changing the algorithm
alvarogi:
About the t-distribution: when the number of samples is small (i.e., at the beginning of the sampling or when the SNR is relatively large), we should use this distribution to estimate the position of the true mean. The size of the confidence interval is then build by multiplying a t-value by the standard error of the mean. If the number of samples is large, the t-value converges to 1 and this is not relevant anymore, since we will be basically using the standard error of the mean.
kwant-bot
@kwant-bot
alvarogi:
About delta: all the tests I ran (different functions and noise distributions) generally worked well for values between 0.1 and 0.5. You can find analytical bounds for the optimal value here. For your number of samples per point I would say that delta=0.2 should probably work, but I suggest to check the bounds first. Also remember that larger delta favors exploration and smaller delta favors exploitation: if you find that the algorithm is excessively resampling existing points, you can increase delta and keep running it.
alvarogi:
Also, I will share the slides in ~presenting in a few minutes.
basnijholt:
I am surprised there aren't more...