v2.0.1

Change from v2.0.0

Bugfixes

• To load a saved history fails (#52, #59)
• To write the last action to policy fails (#57, #58)

Full Changelog: v2.0.0...v2.0.1

v2.0.0

Important Notice

From v2.0.0, PHYSBO is distributed under the Mozilla Public License (MPL) v2.
v2.0.0より、 PHYSBO は Mozilla Public License (MPL) v2 のもとで公開・頒布されます。

This change makes it clear that code importing PHYSBO will not be affected by copyleft. However, if you modify and redistribute PHYSBO itself, copyleft will still apply, and you will need to redistribute it under MPL v2.
これにより、PHYSBO を import するコードについては、コピーレフトの影響を受けないことが明確になりました。一方、PHYSBO そのものを修正して再頒布する場合には、これまでと同様にコピーレフトが働き、MPL v2 の下で再頒布する必要があります。

What's Changed

• Change License to MPLv2 (#55)
• Set the upper bound of numpy version (#56)

Full Changelog: v1.1.1...v2.0.0

v1.1.1

Changed from v1.1.0

• Bug fixes
  • Use numpy<1.20 for installing physbo for python<=3.9 (#49)
    • This solves the problem of mismatches of the size of numpy objects
  • get_post_fmean, get_post_fcov, and get_score now work just after policy.write (interactive mode) (#42, #51)
  • Remove evaluated actions from loaded policy (#44, #45)
  • Remove actions on writing history (policy.write) (#43, #46)
    • Thanks @k-kitai for making a PR!

Full Changelog: v1.1.0...v1.1.1

v1.1.0

Changed from v1.0.1

• New feature
  • Now history stores elapsed times (#28, #40)
    • time_total: total time in steps
    • time_update_predictor: time for updating model, e.g., learning hyper parameters
    • time_get_action: time for suggesting next actions (candidates)
    • time_run_simulator: time for evaluating objective functions
  • is_disp argument of bayes_search now can suppress the message in the hyper-parameter learning process (#41)
    • NOTE: Previously, users had to set config.learning.is_disp = False in order to suppress.
• Bug fixes
  • Algorithm of multiple-probe search in multiple-objective optimization is fixed (#36)
  • base_predictor.save and base_predictor.load now work (#37)
  • show_start_message_multi_search_mo are replaced by show_start_message_multi (#35, #39)
  • Make self.actions an instance of np.ndarray (#32, #34)

Full Changelog: v1.0.1...v1.1.0

v1.0.1

Release notes

Changes from v1.0.0

• fixed bugs
  • When no predictor is set, get_score, get_post_fmean, and get_post_fcov automatically initialize it by gp.predictor (#18, #19)
  • discrete.policy.get_score("TS") returns acquisition function as a vector, instead of a matrix (#19)
• documents
  • added new examples in the examples directory (#27)

v1.0.0

Release note

Changes from v0.3.x

New features

• policy.get_post_fmean(xs) (#17)
  • mean values of the trained predictor (the post-distribution of the Guassian process) at points xs
• policy.get_post_fcov(xs) (#17)
  • covariance of the trained predictor (the post-distribution of the Guassian process) at points xs
• policy.get_score(mode, xs_or_actions) (#16)
  • score (acquisition function) at points xs or actions

Changes

• Move physbo.search.discrete.policy_mo to physbo.search.discrete_multi.policy (#13)

Fixes

• Return best actions as an array of integer (#12)
• Fixed a bug of crashing if no actions remain (#14)

Documents

• English manual is uploaded.

v0.3.0

Release note

Changes from v0.2.x

New features

• Multi-objective optimization (Pareto optimization)
• Initialize policy (model) with pre-evaluated training datasets
• Parallelization for evaluating acquisition function (score) on each candidate (EXPERIMENTAL)

Documents

• Tutorials are updated
  • Multi-objective optimization
  • Initialize with pre-evaluated data

Others

• pip install from the local source code is enabled

v0.2.0

Release Note

Change from v0.1.x

• 🎉 Support Python3 (>= 3.6)
  • No longer support Python2 (Use v0.1.0)
• Manual, README, and sample files are updated.

v0.1.0

optimization tools for PHYsics based on Bayesian Optimization ( PHYSBO )

Bayesian optimization has been proven as an effective tool in accelerating scientific discovery.
A standard implementation (e.g., scikit-learn), however, can accommodate only small training data.
PHYSBO is highly scalable due to an efficient protocol that employs Thompson sampling, random feature maps, one-rank Cholesky update and automatic hyperparameter tuning. Technical features are described in COMBO's document.
PHYSBO was developed based on COMBO for academic use.

Document

• english (in preparation)
• 日本語

Required Packages

• Python 2.7.x
  • We plan to support Python 3.x in the next version of PHYSBO
• numpy
• scipy

Install

• From PyPI (recommended)

    $ pip2 install physbo

• From source (for developers)

  1. Install NumPy and Cython before installing PHYSBO

     $ pip2 install numpy Cython

  2. Download or clone the github repository

     $ git clone https://github.com/issp-center-dev/PHYSBO
     

  3. Run setup.py install

     $ cd physbo
     $ python2 setup.py install --user

  4. Note: Do not import physbo at the root directory of the repository because import physbo does not try to import the installed PHYSBO but one in the repository, which includes Cython codes not compiled.

Uninstall

$ pip2 uninstall physbo

Usage

After installation, you can launch the test suite from 'examples/grain_bound/tutorial.ipynb'.

License

PHYSBO was developed based on COMBO for academic use.
This package is distributed under GNU General Public License version 3 (GPL v3) or later.

Copyright

© 2020- The University of Tokyo. All rights reserved.
This software was developed with the support of "Project for advancement of software usability in materials science" of The Institute for Solid State Physics, The University of Tokyo.