Notebook Entry | moorepants

Tue 23 September 2014
notebook
Jason K. Moore
#notebook, #aio

AIO for September 23, 2014

Last week's accomplishments

Objectives from last week

Lab Stuff

[6 hr, took 9.6 hr] Create a presentation for the lab seminar on the 17th. -- Done: http://www.moorepants.info/presentations/2014/csu-fall-hmc-seminar/fall-csu-hmc-seminar.slides.html

Perturbed Walking Data Paper

[1 hr, took 1.6 hr] Write draft of data processing software section. -- Done: https://github.com/csu-hmc/perturbed-data-paper/commit/207f679949114d18701829538e6352ad0f944088

Parameter Identification with Direct Collocation Paper

[1 hr, took 1.2 hr] Add a section about how the direct collocation is implemented with a backward Euler integration assumption. -- Done: https://github.com/csu-hmc/inverted-pendulum-sys-id-paper/commit/40a78b939cd51adfa19c6be367faf31525e4894f

Parameter Identification with Direct Collocation Source Code

Spent 11 hrs on this stuff instead of 3...

[2 hr] Add the ability to specify constants and arrays explicitly in ufuncify_matrix to avoid creating arrays of constants in the jacobian evaluation function. -- Done: https://github.com/csu-hmc/inverted-pendulum-sys-id/commit/8502a154914d6345158c92ffaac1fb5d3946a532
[1 hr] Move the indice computation out of the jacobian evaluation computation. -- Done: https://github.com/csu-hmc/inverted-pendulum-sys-id/commit/ee87327a564c736e4eaa375097e72e689f5638ff

With these final changes the speed of the constraint function evaluation is really fast. For example, when comparing to the original naive implementation of the constraint Jacobian evaluation I get a ~110X speedup for the case below where there are 1,056,000 non-zero values in the Jacobian. The function evaluations are now ~10X faster than the IPOPT computation time.

The new (7094e45a42):

In [1]: run pendulum.py -n 1 -d 480.0 -r -s 100.0 -i known

In [2]: %timeit identifier.con_jac_func(identifier.initial_guess)
1 loops, best of 3: 126 ms per loop

In [3]: len(identifier.initial_guess)
Out[3]: 192008

The old (bc00f1d):

In [1]: run pendulum.py -n 1 -d 480.0 -r -s 100.0 -i known

In [3]: %timeit identifier.con_jac_func(identifier.initial_guess)
1 loops, best of 3: 13.9 s per loop

In [4]: 13.9 / .126
Out[4]: 110.31746031746032

Bicycle Steer Torque Paper

Spent 1.9 hrs on this, most on improving the plots...see issues.

[0.5 hr] Add in a good photo of the actual steer torque design to go along with the CAD model. -- Done: https://github.com/moorepants/steer-torque-manuscript/commit/987ae39dca43a4234666625833adf24bfb39b81c

Yeadon Paper

[0.25 hr, took 0.6 hr] Commit latest draft from the journal to Git. -- Done: https://github.com/chrisdembia/python-yeadon-paper/commit/88523bac88325eec74a56ef661442c9897eb1ba5

Other things I did

Yeadon paper was published and is now waiting peer review: http://f1000research.com/articles/3-223/v1
Implemented a temp directory for ufuncify_matrix code gen files: https://github.com/csu-hmc/inverted-pendulum-sys-id/commit/7094e45a42b1334bf066beac1f4c9b63fd6a9fec
Read a nice paper by Victoria Stodden on reproducibility: http://web.stanford.edu/~vcs/papers/SMPRCS2010.pdf

Last week's issues

I really overestimated my time, especially since I didn't work Friday, Saturday, or Sunday. So there are a bunch of tasks I didn't finish. I also spent too much time on the tasks I did do and wasted time learning about general collocation and optimization tools and working on my general implementation with SymPy.

Lab Stuff

[0.5 hr] Make sure that the hmc.csuohio.edu offsite backup can be restored.

Bicycle Steer Torque Paper

[1 hr] Change R plots to a prettier ggplot2 version. -- This is mostly done. I was having some trouble with the ggplot histogram and also need to get the fonts to match the fonts in the paper.
[1 hr] Improve the matplotlib plots to make them prettier.
[0.5 hr] Figure out the best citation for Kageyama's design.

Bicycle Data Paper

[1 hr] Finish raw data README.
[2 hr] Create calibration data README.
[0.5 hr] Try to upload the davis instrumented bicycle h5 file (with raw and processed data) to Figshare.

Job Search

[2 hours] Draft a plan/outline for a new research statement for prof applications.

Software Carpentry

[1 hr] Respond to Nick Hermez.
[1 hr] Decide on a date, confer with Ahmet, and let Software Carpentry know.
[0.5 hr] Make a list of tasks that need to be done so I can divvy them up.

This week's objectives

Lab Stuff

[0.5 hr] Make sure that the hmc.csuohio.edu offsite backup can be restored.
[0.5 hr] Undo the review workflow on the website.

Inverted Pendulum Parameter Estimation Source

[2 hr] Implement both process and measurement noise.
[2 hr] Implement a validation simulation on a portion of the data.
[3 hr] Setup runs to create the convergence and global minima chance plots.
[3 hr] Flesh out the collocator class design more. Specifically how to add extra constraints and maybe even the objective.

Inverted Pendulum Parameter Estimation Paper

[3 hr] Find and read papers on using DC for parameter estimation.
[1 hr] Make a pass through the outline to define things better, i.e. like specific plots, etc.

Perturbed Walking Data Paper

[2 hr] Decide on what will actually be the processed data and write script to produce it from raw.
[3 hr] Add in the belt acceleration compensation to GTK.

Bicycle Steer Torque Paper

[1 hr] Change R plots to a prettier ggplot2 version.
[1.5 hr] Improve the matplotlib plots to make them prettier.
[1 hr] Figure out the best citation for Kageyama's design.
[0.25 hr] Check in with Mont about his review at end of week.

Bicycle Data Paper

[1 hr] Finish raw data README.
[2 hr] Create calibration data README.
[1 hr] Try to upload the davis instrumented bicycle h5 file (with raw and processed data) to Figshare.

Job Search

[2 hours] Draft a plan/outline for a new research statement for prof applications.

Software Carpentry

[0.5 hr] Respond to Nick Hermez.
[1 hr] Decide on a date, confer with Ahmet, and let Software Carpentry know.
[0.5 hr] Make a list of tasks that need to be done so I can divvy them up.