Merge pull request #1517 from pints-team/xnes-cov-matrix-update-fix

Fixed bug in XNES covariance matrix update.

Author: MichaelClerx

CHANGELOG.md

@@ -5,13 +5,14 @@ All notable changes to this project will be documented in this file.
 ## Unreleased
 
 ### Added
-- [#1499](https://github.com/pints-team/pints/pull/1499) Added a log-uniform prior class.
 - [#1505](https://github.com/pints-team/pints/pull/1505) Added notes to `ErrorMeasure` and `LogPDF` to say parameters must be real and continuous.
+- [#1499](https://github.com/pints-team/pints/pull/1499) Added a log-uniform prior class.
 ### Changed
 - [#1503](https://github.com/pints-team/pints/pull/1503) Stopped showing time units in controller logs, because the units change depending on the output type (see #1467).
 ### Deprecated
 ### Removed
 ### Fixed
+- [#1517](https://github.com/pints-team/pints/pull/1517) Fixed a major bug in the covariance matrix update for xNES.
 - [#1505](https://github.com/pints-team/pints/pull/1505) Fixed issues with toy problems that accept invalid inputs.
 
 

pints/_optimisers/_xnes.py

@@ -32,6 +32,8 @@ class XNES(pints.PopulationBasedOptimiser):
 
     .. [2] PyBrain: The Python machine learning library
            http://pybrain.org
+           PyBrain is co-authored by xNES' authors.
+
     """
     def __init__(self, x0, sigma0=None, boundaries=None):
         super(XNES, self).__init__(x0, sigma0, boundaries)
@@ -47,7 +49,7 @@ def __init__(self, x0, sigma0=None, boundaries=None):
         self._bounded_ids = None  # Indices of those xs
 
         # Normalisation / distribution
-        self._mu = np.array(self._x0)   # Mean
+        self._mu = pints.vector(x0)     # Mean
         self._A = None                  # Covariance
 
         # Best solution seen
@@ -106,13 +108,13 @@ def _initialise(self):
         d = self._n_parameters
         n = self._population_size
 
-        # Learning rates
+        # Learning rates, see Table 1 in [1]
         # TODO Allow changing before run() with method call
         self._eta_mu = 1
         # TODO Allow changing before run() with method call
         self._eta_A = 0.6 * (3 + np.log(d)) * d ** -1.5
 
-        # Pre-calculated utilities
+        # Pre-calculated utilities, see Table 1 in [1]
         self._us = np.maximum(0, np.log(n / 2 + 1) - np.log(1 + np.arange(n)))
         self._us /= np.sum(self._us)
         self._us -= 1 / n
@@ -162,10 +164,12 @@ def tell(self, fx):
         self._mu += self._eta_mu * np.dot(self._A, Gd)
 
         # Update root of covariance matrix
+        # Note that this is equation 11 (for the eta-sigma=eta-B case), not the
+        # more general equations 9&10 version given in Algorithm 1
         Gm = np.dot(
             np.array([np.outer(z, z).T - self._I for z in self._zs]).T,
             self._us)
-        self._A *= scipy.linalg.expm(np.dot(0.5 * self._eta_A, Gm))
+        self._A = np.dot(self._A, scipy.linalg.expm(0.5 * self._eta_A * Gm))
 
         # Update f_guessed on the assumption that the lowest value in our
         # sample approximates f(mu)

pints/tests/test_opt_controller.py

@@ -233,6 +233,7 @@ def test_stopping_max_iterations(self):
     def test_logging(self):
 
         # Test with logpdf
+        np.random.seed(1)
         r = pints.toy.TwistedGaussianLogPDF(2, 0.01)
         x = np.array([0, 1.01])
         b = pints.RectangularBoundaries([-0.01, 0.95], [0.01, 1.05])
@@ -255,17 +256,17 @@ def test_logging(self):
         self.assertEqual(lines[5][:-3],
                          '0     3     -4.140462 -4.140462   0:0')
         self.assertEqual(lines[6][:-3],
-                         '1     6     -4.140462 -4.140465   0:0')
+                         '1     6     -4.140462 -4.140482   0:0')
         self.assertEqual(lines[7][:-3],
-                         '2     11    -4.140462 -4.140462   0:0')
+                         '2     9     -4.140462 -4.140465   0:0')
         self.assertEqual(lines[8][:-3],
-                         '3     16    -4.140462 -4.140466   0:0')
+                         '3     14    -4.140462 -4.140462   0:0')
         self.assertEqual(lines[9][:-3],
-                         '6     33    -4.140462 -4.140462   0:0')
+                         '6     30    -4.140462 -4.140462   0:0')
         self.assertEqual(lines[10][:-3],
-                         '9     51    -4.140462 -4.140462   0:0')
+                         '9     47    -4.140462 -4.140463   0:0')
         self.assertEqual(lines[11][:-3],
-                         '10    51    -4.140462 -4.140462   0:0')
+                         '10    47    -4.140462 -4.140463   0:0')
         self.assertEqual(
             lines[12], 'Halting: Maximum number of iterations (10) reached.')
 
@@ -448,8 +449,8 @@ def test_post_run_statistics(self):
         opt.run()
         t_upper = t.time()
 
-        self.assertEqual(opt.iterations(), 84)
-        self.assertEqual(opt.evaluations(), 495)
+        self.assertEqual(opt.iterations(), 125)
+        self.assertEqual(opt.evaluations(), 734)
 
         # Time after run is greater than zero
         self.assertIsInstance(opt.time(), float)