fgerzer / apsis Goto Github PK

At the moment we optimize EI with scypy.optimize.minimize. Tried several optimizer methods, all of them get stuck in local extrema.

Problem seems to be well known:

• the bergstra group reports on it http://papers.nips.cc/paper/4443-algorithms-for-hyper-parameter-optimization.pdf (page 3)
• http://www.icml2010.org/papers/297.pdf report about this, too
• the old spearmint package also seems to use grid search here if I understand their code right

Proposed optimization methods there

• grid search (probably not feasible)
• evolutionary algorithm

Other ideas

• random search, how many points?
• continue using SLSQP, L-BFGS-B and implement a random restart for them. How often?

Make them more concise; inline comments, more tests and cleaner structure.

Use lower/upper bound in constructor, or use warping function.

When generating five random points, the gp seems to base it only on the last four.

After the initial random points, only adds 0.5 as a new point.
Corresponding warning seems to be

/usr/local/lib/python2.7/dist-packages/GPy/core/transformations.py:32: RuntimeWarning: overflow encountered in exp
  return np.where(x>lim_val, x, np.log(1. + np.exp(x)))
/usr/local/lib/python2.7/dist-packages/GPy/kern/parts/rbf.py:256: RuntimeWarning: divide by zero encountered in divide
  X = X / self.lengthscale
/usr/local/lib/python2.7/dist-packages/GPy/kern/parts/rbf.py:258: RuntimeWarning: invalid value encountered in add
  self._K_dist2 = -2.*tdot(X) + (Xsquare[:, None] + Xsquare[None, :])
/usr/local/lib/python2.7/dist-packages/GPy/util/linalg.py:121: RuntimeWarning: invalid value encountered in less_equal
  if np.any(diagA <= 0.):

Like with f259835

to make sure there is consistency between eq and hash.

Implement a test framework on pre-computed grid points.
Should also include comparisons between cores, plots and similar things.

This means that we can transfer all non-specific optimization code to the general AcqFunction, and just add specific (maybe with an "analytical" keyword) compute_max functions to some AcqFunctions.

See Justin's mail, 2014-11-10:

"Vorneweg: ich faends cool wenn wir gitter [1] benutzen wuerden. Das ist ein persistent chat fuer github repos... das waere evt effizient. Allerdings muss Frederik das glaube ich aufmachen."

• write tests to test the compare_values() method

Add missing docu, p.ex. the init.

Document testCandidate.

Check implementation of EI.

-> There might be a mistake using variance and std dev interchanged.
-> make it multi dimensional

candidate lists in RandomSearchCore are shared among all class instances => create instance in constructor to fix

=> from justins mail:

https://github.com/FrederikDiehl/BayOpt/blob/master/code/apsis/apsis/RandomSearchCore.py#L19

Da steckt nen fieses Problem drin. Und zwar, wird hier finished_candidates zu einer Klassenvariable. Wenn man da Appends drauf macht, so wird das bei jedem Objekt dieser Klasse sichtbar, weil die sich das alles teilen. Das ist jetzt nicht sonderlich wichtig im Betrieb, da ja in der Regel nur ein Objekt dieser Klasse pro Prozess laeuft. Beim Unittesten kann einen das aber ziemlich beissen und dann sucht man ne ganze Weile...

Loesung:

class blubb(object):

candidates = None

def init(self, ...):
self.candidates = []

Init currently assumes four values (lower_bound, upper_bound, minimization_problem and random_state) all exist in the param dict. Add default values and/or errors.

Investigate performance and see better implementations of gaussian cdfs. Needed in AcquisitionFunctions, at the moment, Probability of Improvement.

Candidate requires class documentation.

Is missing documentation except for next_candidate.

Candidate._init currently does not check parameter validity.

From Justin's comment on #6:

I encourage you to stick to 79 characters per line.

There are many reasons for this, e.g.

human eye is much better at reading short lines (that's why newspapers do it like that),
less instructions contained in one line, which is also the atom in version control systems
possibility to view source code side by side

Currently, SimpleBayesianOptimizationCore returns the same point for each worker that asks for one. This will have to be changed.
This also includes implementing working timeouts.

In RandomSearchCore.working, we are currently only checking whether a candidate is in the list, but does not check equality in points.

Is currently completely undocumented.

Write a small tutorial on how to use apsis.

Install automated documentation system.

Class needs documentation; so does is_better_candidate_as.