Merge pull request #188 from python-adaptive/cycle_strategy

BalancingLearner: add a "cycle" strategy, sampling the learners one by one

Author: basnijholt

adaptive/learner/balancing_learner.py

@@ -4,6 +4,7 @@
 from collections.abc import Iterable
 from contextlib import suppress
 from functools import partial
+import itertools
 from operator import itemgetter
 
 import numpy as np
@@ -50,10 +51,11 @@ class BalancingLearner(BaseLearner):
     function : callable
         A function that calls the functions of the underlying learners.
         Its signature is ``function(learner_index, point)``.
-    strategy : 'loss_improvements' (default), 'loss', or 'npoints'
+    strategy : 'loss_improvements' (default), 'loss', 'npoints', or 'cycle'.
         The points that the `BalancingLearner` choses can be either based on:
         the best 'loss_improvements', the smallest total 'loss' of the
-        child learners, or the number of points per learner, using 'npoints'.
+        child learners, the number of points per learner, using 'npoints',
+        or by cycling through the learners one by one using 'cycle'.
         One can dynamically change the strategy while the simulation is
         running by changing the ``learner.strategy`` attribute.
 
@@ -90,10 +92,11 @@ def __init__(self, learners, *, cdims=None, strategy="loss_improvements"):
 
     @property
     def strategy(self):
-        """Can be either 'loss_improvements' (default), 'loss', or 'npoints'
-        The points that the `BalancingLearner` choses can be either based on:
-        the best 'loss_improvements', the smallest total 'loss' of the
-        child learners, or the number of points per learner, using 'npoints'.
+        """Can be either 'loss_improvements' (default), 'loss', 'npoints', or
+        'cycle'. The points that the `BalancingLearner` choses can be either
+        based on: the best 'loss_improvements', the smallest total 'loss' of
+        the child learners, the number of points per learner, using 'npoints',
+        or by going through all learners one by one using 'cycle'.
         One can dynamically change the strategy while the simulation is
         running by changing the ``learner.strategy`` attribute."""
         return self._strategy
@@ -107,10 +110,13 @@ def strategy(self, strategy):
             self._ask_and_tell = self._ask_and_tell_based_on_loss
         elif strategy == "npoints":
             self._ask_and_tell = self._ask_and_tell_based_on_npoints
+        elif strategy == "cycle":
+            self._ask_and_tell = self._ask_and_tell_based_on_cycle
+            self._cycle = itertools.cycle(range(len(self.learners)))
         else:
             raise ValueError(
-                'Only strategy="loss_improvements", strategy="loss", or'
-                ' strategy="npoints" is implemented.'
+                'Only strategy="loss_improvements", strategy="loss",'
+                ' strategy="npoints", or strategy="cycle" is implemented.'
             )
 
     def _ask_and_tell_based_on_loss_improvements(self, n):
@@ -173,6 +179,17 @@ def _ask_and_tell_based_on_npoints(self, n):
         points, loss_improvements = map(list, zip(*selected))
         return points, loss_improvements
 
+    def _ask_and_tell_based_on_cycle(self, n):
+        points, loss_improvements = [], []
+        for _ in range(n):
+            index = next(self._cycle)
+            point, loss_improvement = self.learners[index].ask(n=1)
+            points.append((index, point[0]))
+            loss_improvements.append(loss_improvement[0])
+            self.tell_pending((index, point[0]))
+
+        return points, loss_improvements
+
     def ask(self, n, tell_pending=True):
         """Chose points for learners."""
         if n == 0:

adaptive/tests/test_balancing_learner.py

@@ -6,6 +6,9 @@
 from adaptive.runner import simple
 
 
+strategies = ["loss", "loss_improvements", "npoints", "cycle"]
+
+
 def test_balancing_learner_loss_cache():
     learner = Learner1D(lambda x: x, bounds=(-1, 1))
     learner.tell(-1, -1)
@@ -26,7 +29,7 @@ def test_balancing_learner_loss_cache():
     assert bl.loss(real=True) == real_loss
 
 
-@pytest.mark.parametrize("strategy", ["loss", "loss_improvements", "npoints"])
+@pytest.mark.parametrize("strategy", strategies)
 def test_distribute_first_points_over_learners(strategy):
     for initial_points in [0, 3]:
         learners = [Learner1D(lambda x: x, bounds=(-1, 1)) for i in range(10)]
@@ -41,7 +44,7 @@ def test_distribute_first_points_over_learners(strategy):
         assert len(set(i_learner)) == len(learners)
 
 
-@pytest.mark.parametrize("strategy", ["loss", "loss_improvements", "npoints"])
+@pytest.mark.parametrize("strategy", strategies)
 def test_ask_0(strategy):
     learners = [Learner1D(lambda x: x, bounds=(-1, 1)) for i in range(10)]
     learner = BalancingLearner(learners, strategy=strategy)
@@ -55,6 +58,7 @@ def test_ask_0(strategy):
         ("loss", lambda l: l.loss() < 0.1),
         ("loss_improvements", lambda l: l.loss() < 0.1),
         ("npoints", lambda bl: all(l.npoints > 10 for l in bl.learners)),
+        ("cycle", lambda l: l.loss() < 0.1),
     ],
 )
 def test_strategies(strategy, goal):