New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Added a simple particle swarm based hyper-heuristic #125

Open

Alan-Robertson wants to merge 16 commits into QInfer:feature-hho from Alan-Robertson:feature-hyper-heuristic-optimisation

Alan-Robertson commented

•

First hyper-heuristic optimisation algorithm to be added for this pull request, an example follows:
First we define a test function

f_rosenbrock = lambda x: numpy.sum([((x[i+1]  - x[i]**2)**2 + (1 - x[i])**2)/len(x) for i in range(len(x)-1)])

Next we initialise the hyper-heuristic optimiser and pass it the relevant parameters over which to optimise before calling it to perform the optimisation

hh_opt = ParticleSwarmOptimiser(['x','y'], fitness_function = f_rosenbrock)
hh_opt()

Similarly this is parrallelisable given a map function

import numpy
from ipyparallel import Client
from functools import partial
import pickle as pkl
import dill

f_rosenbrock = lambda x: numpy.sum([((x[i+1]  - x[i]**2)**2 + (1 - x[i])**2)/len(x) for i in range(len(x)-1)])

c = Client(profile='testing_profile')
c[:].use_dill()
dview = c[:]

hh_opt = ParticleSwarmOptimiser(['x','y','z','a'], fitness_function = f_rosenbrock)
hh_opt(serial_map=dview.map_sync)


          Added a simple particle swarm based hyper-heuristic

82fe4cd

coveralls commented

•

Coverage decreased (-1.3%) to 70.407% when pulling 82fe4cd on Alan-Robertson:feature-hyper-heuristic-optimisation into 8118ab7 on QInfer:master.

cgranade requested changes

View reviewed changes

Collaborator

cgranade left a comment

Thank you for the submission, @Alan-Robertson! It's looking like a really good feature so far, so thank you again. I've left some comments on some individual issues, but I think it's looking really good. In combination with some tests and some documentation, I think this will be awesome to have included. ♥

src/qinfer/hyper_heuristic_optimisers.py Outdated

		@@ -0,0 +1,202 @@

Collaborator

cgranade

We should make sure the copyright header and -*- coding lines are kept on this file as well.

src/qinfer/hyper_heuristic_optimisers.py Outdated

+              ## CLASSES ####################################################################
+              __all__ = [
+                  'ParticleSwarmOptimiser'

Collaborator

cgranade

The performance objective class should also be listed under __all__.

src/qinfer/hyper_heuristic_optimisers.py Outdated

+                  'ParticleSwarmOptimiser'
+              ]
+              class HyperHeuristicOptimiser(object):

Collaborator

cgranade

At the level of generality that this class is written, including HyperHeuristic in the name may be slightly misleading? I think it makes a lot of sense to keep the default choice of function being the performance objective below, but to rename here as a nod to that generality. Perhaps even just Optimiser (or even Optimizer, since the project was started in American English instead of something a bit more common)...

src/qinfer/hyper_heuristic_optimisers.py Outdated

+                      self._funct_kwargs = funct_kwargs
+                      if fitness_function is None: # Default to calling perf test multiple
+                          self._optimisable = PerfTestMultipleAbstractor(

Collaborator

cgranade

Should this be _optimisable or _fitness_function? I don't see any other references to _optimisable?

src/qinfer/hyper_heuristic_optimisers.py Outdated

+                  def fitness_function(self, params):
+                      return self._fitness_function(params)
+                  def parrallel(self):

Collaborator

cgranade

I think there's only one "r" in parallel?

src/qinfer/hyper_heuristic_optimisers.py Outdated

+                      omega_v=0.35,
+                      phi_p=0.25,
+                      phi_g=0.5,
+                      serial_map=map

Collaborator

cgranade

The name serial_map is a bit misleading here, since it's either serial or parallel depending?

src/qinfer/hyper_heuristic_optimisers.py Outdated

+                              self._fitness[itr - 1]["velocities"])
+                          # Apply the boundary conditions if any exist
+                          if self._boundary_map is not None:

Collaborator

cgranade

The behavior of _boundary_map seems to be similar to Model.canonicalize rather than Models.are_models_valid? If so, that's quite reasonable, but should be clearly indicated in the documentation since it's an unusual way to specify boundaries.

src/qinfer/hyper_heuristic_optimisers.py Outdated

+                      fitness = serial_map(self.fitness_function, particles)
+                      return fitness
+                  def update_positions(self, positions, velocities):

Collaborator

cgranade

These methods are public and should thus have at least some brief docstrings. They probably don't need to be very detailed, but there should be something.

src/qinfer/hyper_heuristic_optimisers.py Outdated

+                               *args,
+                               **kwargs):
+                      self._heuristic = kwargs['heuristic_class']
+                      del kwargs['heuristic_class']

Collaborator

cgranade

This can cause an exception if heuristic_class isn't defined. That's clearly an error, since this class doesn't make any sense without a heuristic class, but the exception that gets raised isn't immediately obvious. It might make sense to promote heuristic_class to being an explicitly named kwarg so that exceptions raised are more readable.

src/qinfer/hyper_heuristic_optimisers.py Outdated

+                      global_attractor = local_attractors[np.argmin(local_attractors["fitness"])]
+                      return local_attractors, global_attractor
+              class PerfTestMultipleAbstractor:

Collaborator

cgranade

I think the term Abstractor here doesn't quite communicate what the class is doing. I'd suggest something like PerformanceObjective, perhaps? Also, this should inherit from object so that it's forced to be a new-style class even on Python 2.

Collaborator

cgranade commented

It seems that unrelated to the PR itself, there's build failures that are caused by Travis not being properly configured for QuTiP 4.1 (see discussion at #126). I'll work on getting that diagnosed, then, so we can include this PR ASAP.

cgranade mentioned this pull request

Update travis config for QuTiP 4.1. #127

Merged

Collaborator

cgranade commented

When you get a second, can you pull master into this PR to include #127? That should help fix build errors. Thanks!


          Merge branch 'master' of github.com:QInfer/python-qinfer into feature…

…-hyper-heuristic-optimisation

coveralls commented

•

Coverage decreased (-1.4%) to 70.407% when pulling 0325109 on Alan-Robertson:feature-hyper-heuristic-optimisation into 75f8723 on QInfer:master.

Collaborator

cgranade commented

It looks like pulling in master helped fix the issues with the build, thanks again for that!

Contributor

taalexander commented

I think some documentation on use cases for the hyper-heuristic optimiser would be nice. From this commit alone I am having trouble piecing it together. I assume this is related to the paper recently listed on the Arxiv by Alan, and Chris but I have yet to find time to read it.

cgranade added the enhancement label

cgranade assigned cgranade and Alan-Robertson

cgranade added the wip label

Collaborator

cgranade commented

I agree entirely, the plan is to add documentation to both the guide and apiref, including descriptions of a few explicit usecases. At the moment, we wanted to start the PR as-is to get initial feedback on the design, since ideally the particle swarm optimizer will be useful outside the immediate usecases that @Alan-Robertson described in his initial comment.


          Applied edits with regards to comments provided

60ce342

coveralls commented

•

Coverage decreased (-1.4%) to 70.407% when pulling 60ce342 on Alan-Robertson:feature-hyper-heuristic-optimisation into 75f8723 on QInfer:master.


          Moved from map to apply, along with other changes

a2762ec

coveralls commented

•

Coverage decreased (-0.7%) to 71.108% when pulling a2762ec on Alan-Robertson:feature-hyper-heuristic-optimisation into 75f8723 on QInfer:master.

Alan-Robertson and others added 7 commits

April 19, 2017 13:56


          Added tests and additional PSO variants

8d35435


          Merge branch 'master' of github.com:QInfer/python-qinfer into feature…

fdafb22

…-hyper-heuristic-optimisation


          Spelling swapped to US, renamed some classes, removed long testing fo…

0f1f598

…r travis compatibility


          Fixed import, tabs → spaces.

95094e6


          PEP8 / pylint fixes

339fc5d


          Import fixes, some tests now pass.

60f3874


          More import fixes, refactored to test annealing+tempering.

cab23a9

cgranade mentioned this pull request

Fix hho tests Alan-Robertson/python-qinfer#1

Merged

Alan-Robertson added 2 commits

April 20, 2017 10:51


          Merge pull request #1 from cgranade/fix-hho-tests

d9964ff

Fix hho tests


          Added SPSA optimiser/optimizer

9dd8c12

cgranade changed the base branch from master to feature-hho

May 12, 2017 06:05

Alan-Robertson added 3 commits

May 18, 2017 14:39


          Fixed apply issue for python 3

73875a3


          Fixed issue with boundary map

b87a69c


          Fixed issue with boundary map

58dbe05

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

enhancement wip