Unit tests for the :class:`SymmetricLinearGame` class.
"""
-MAX_COND = 250
-"""
-The maximum condition number of a randomly-generated game.
-"""
-
-RANDOM_MAX = 10
-"""
-When generating uniform random real numbers, this will be used as the
-largest allowed magnitude. It keeps our condition numbers down and other
-properties within reason.
-"""
-
-from math import sqrt
-from random import randint, uniform
from unittest import TestCase
-from cvxopt import matrix
-from dunshire.cones import NonnegativeOrthant, IceCream
from dunshire.games import SymmetricLinearGame
-from dunshire.matrices import (append_col, append_row, eigenvalues_re,
- identity, inner_product)
+from dunshire.matrices import eigenvalues_re, inner_product
from dunshire import options
-
-
-def random_matrix(dims):
- """
- Generate a random square matrix.
-
- Parameters
- ----------
-
- dims : int
- The number of rows/columns you want in the returned matrix.
-
- Returns
- -------
-
- matrix
- A new matrix whose entries are random floats chosen uniformly from
- the interval [-RANDOM_MAX, RANDOM_MAX].
-
- Examples
- --------
-
- >>> A = random_matrix(3)
- >>> A.size
- (3, 3)
-
- """
- return matrix([[uniform(-RANDOM_MAX, RANDOM_MAX) for _ in range(dims)]
- for _ in range(dims)])
-
-
-def random_nonnegative_matrix(dims):
- """
- Generate a random square matrix with nonnegative entries.
-
- Parameters
- ----------
-
- dims : int
- The number of rows/columns you want in the returned matrix.
-
- Returns
- -------
-
- matrix
- A new matrix whose entries are random floats chosen uniformly from
- the interval [0, RANDOM_MAX].
-
- Examples
- --------
-
- >>> A = random_nonnegative_matrix(3)
- >>> A.size
- (3, 3)
- >>> all([entry >= 0 for entry in A])
- True
-
- """
- L = random_matrix(dims)
- return matrix([abs(entry) for entry in L], (dims, dims))
-
-
-def random_diagonal_matrix(dims):
- """
- Generate a random square matrix with zero off-diagonal entries.
-
- These matrices are Lyapunov-like on the nonnegative orthant, as is
- fairly easy to see.
-
- Parameters
- ----------
-
- dims : int
- The number of rows/columns you want in the returned matrix.
-
- Returns
- -------
-
- matrix
- A new matrix whose diagonal entries are random floats chosen
- uniformly from the interval [-RANDOM_MAX, RANDOM_MAX] and whose
- off-diagonal entries are zero.
-
- Examples
- --------
-
- >>> A = random_diagonal_matrix(3)
- >>> A.size
- (3, 3)
- >>> A[0,1] == A[0,2] == A[1,0] == A[2,0] == A[1,2] == A[2,1] == 0
- True
-
- """
- return matrix([[uniform(-RANDOM_MAX, RANDOM_MAX)*int(i == j)
- for i in range(dims)]
- for j in range(dims)])
-
-
-def random_skew_symmetric_matrix(dims):
- """
- Generate a random skew-symmetrix matrix.
-
- Parameters
- ----------
-
- dims : int
- The number of rows/columns you want in the returned matrix.
-
- Returns
- -------
-
- matrix
- A new skew-matrix whose strictly above-diagonal entries are
- random floats chosen uniformly from the interval
- [-RANDOM_MAX, RANDOM_MAX].
-
- Examples
- --------
-
- >>> A = random_skew_symmetric_matrix(3)
- >>> A.size
- (3, 3)
-
- >>> from dunshire.matrices import norm
- >>> A = random_skew_symmetric_matrix(randint(1, 10))
- >>> norm(A + A.trans()) < options.ABS_TOL
- True
-
- """
- strict_ut = [[uniform(-10, 10)*int(i < j) for i in range(dims)]
- for j in range(dims)]
-
- strict_ut = matrix(strict_ut, (dims, dims))
- return strict_ut - strict_ut.trans()
-
-
-def random_lyapunov_like_icecream(dims):
- r"""
- Generate a random matrix Lyapunov-like on the ice-cream cone.
-
- The form of these matrices is cited in Gowda and Tao
- [GowdaTao]_. The scalar ``a`` and the vector ``b`` (using their
- notation) are easy to generate. The submatrix ``D`` is a little
- trickier, but it can be found noticing that :math:`C + C^{T} = 0`
- for a skew-symmetric matrix :math:`C` implying that :math:`C + C^{T}
- + \left(2a\right)I = \left(2a\right)I`. Thus we can stick an
- :math:`aI` with each of :math:`C,C^{T}` and let those be our
- :math:`D,D^{T}`.
-
- Parameters
- ----------
-
- dims : int
- The dimension of the ice-cream cone (not of the matrix you want!)
- on which the returned matrix should be Lyapunov-like.
-
- Returns
- -------
-
- matrix
- A new matrix, Lyapunov-like on the ice-cream cone in ``dims``
- dimensions, whose free entries are random floats chosen uniformly
- from the interval [-10, 10].
-
- References
- ----------
-
- .. [GowdaTao] M. S. Gowda and J. Tao. On the bilinearity rank of a
- proper cone and Lyapunov-like transformations. Mathematical
- Programming, 147:155-170, 2014.
-
- Examples
- --------
-
- >>> L = random_lyapunov_like_icecream(3)
- >>> L.size
- (3, 3)
- >>> x = matrix([1,1,0])
- >>> s = matrix([1,-1,0])
- >>> abs(inner_product(L*x, s)) < options.ABS_TOL
- True
-
- """
- a = matrix([uniform(-10, 10)], (1, 1))
- b = matrix([uniform(-10, 10) for _ in range(dims-1)], (dims-1, 1))
- D = random_skew_symmetric_matrix(dims-1) + a*identity(dims-1)
- row1 = append_col(a, b.trans())
- row2 = append_col(b, D)
- return append_row(row1, row2)
-
-
-def random_orthant_game():
- """
- Generate the ``L``, ``K``, ``e1``, and ``e2`` parameters for a
- random game over the nonnegative orthant, and return the
- corresponding :class:`SymmetricLinearGame`.
-
- We keep going until we generate a game with a condition number under
- 5000.
- """
- ambient_dim = randint(1, 10)
- K = NonnegativeOrthant(ambient_dim)
- e1 = [uniform(0.5, 10) for _ in range(K.dimension())]
- e2 = [uniform(0.5, 10) for _ in range(K.dimension())]
- L = random_matrix(K.dimension())
- G = SymmetricLinearGame(L, K, e1, e2)
-
- if G._condition() <= MAX_COND:
- return G
- else:
- return random_orthant_game()
-
-
-def random_icecream_game():
- """
- Generate the ``L``, ``K``, ``e1``, and ``e2`` parameters for a
- random game over the ice-cream cone, and return the corresponding
- :class:`SymmetricLinearGame`.
- """
- # Use a minimum dimension of two to avoid divide-by-zero in
- # the fudge factor we make up later.
- ambient_dim = randint(2, 10)
- K = IceCream(ambient_dim)
- e1 = [1] # Set the "height" of e1 to one
- e2 = [1] # And the same for e2
-
- # If we choose the rest of the components of e1,e2 randomly
- # between 0 and 1, then the largest the squared norm of the
- # non-height part of e1,e2 could be is the 1*(dim(K) - 1). We
- # need to make it less than one (the height of the cone) so
- # that the whole thing is in the cone. The norm of the
- # non-height part is sqrt(dim(K) - 1), and we can divide by
- # twice that.
- fudge_factor = 1.0 / (2.0*sqrt(K.dimension() - 1.0))
- e1 += [fudge_factor*uniform(0, 1) for _ in range(K.dimension() - 1)]
- e2 += [fudge_factor*uniform(0, 1) for _ in range(K.dimension() - 1)]
- L = random_matrix(K.dimension())
- G = SymmetricLinearGame(L, K, e1, e2)
-
- if G._condition() <= MAX_COND:
- return G
- else:
- return random_icecream_game()
+from .randomgen import (random_icecream_game, random_ll_icecream_game,
+ random_ll_orthant_game, random_nn_scaling,
+ random_orthant_game, random_positive_orthant_game,
+ random_translation)
# Tell pylint to shut up about the large number of methods.
"""
Tests for the SymmetricLinearGame and Solution classes.
"""
- def assert_within_tol(self, first, second):
+ def assert_within_tol(self, first, second, modifier=1):
"""
Test that ``first`` and ``second`` are equal within a multiple of
our default tolerances.
- The factor of two is because if we compare two solutions, both
- of which may be off by ``ABS_TOL``, then the result could be off
- by ``2*ABS_TOL``. The factor of ``RANDOM_MAX`` allows for
- scaling a result (by ``RANDOM_MAX``) that may be off by
- ``ABS_TOL``. The final factor of two is to allow for the edge
- cases where we get an "unknown" result and need to lower the
- CVXOPT tolerance by a factor of two.
- """
- self.assertTrue(abs(first - second) < 2*2*RANDOM_MAX*options.ABS_TOL)
+ Parameters
+ ----------
+ first : float
+ The first number to compare.
- def assert_solution_exists(self, G):
- """
- Given a SymmetricLinearGame, ensure that it has a solution.
- """
- soln = G.solution()
+ second : float
+ The second number to compare.
- expected = inner_product(G._L*soln.player1_optimal(),
- soln.player2_optimal())
- self.assert_within_tol(soln.game_value(), expected)
+ modifier : float
+ A scaling factor (default: 1) applied to the default
+ tolerance for this comparison. If you have a poorly-
+ conditioned matrix, for example, you may want to set this
+ greater than one.
+
+ """
+ self.assertTrue(abs(first - second) < options.ABS_TOL*modifier)
to a smaller nonnegative real number.
"""
G = random_orthant_game()
- self.assertTrue(G._condition() >= 1.0)
- G = random_icecream_game()
- self.assertTrue(G._condition() >= 1.0)
-
-
- def test_solution_exists_orthant(self):
- """
- Every linear game has a solution, so we should be able to solve
- every symmetric linear game over the NonnegativeOrthant. Pick
- some parameters randomly and give it a shot. The resulting
- optimal solutions should give us the optimal game value when we
- apply the payoff operator to them.
- """
- G = random_orthant_game()
- self.assert_solution_exists(G)
-
-
- def test_solution_exists_icecream(self):
- """
- Like :meth:`test_solution_exists_nonnegative_orthant`, except
- over the ice cream cone.
- """
+ self.assertTrue(G.condition() >= 1.0)
G = random_icecream_game()
- self.assert_solution_exists(G)
+ self.assertTrue(G.condition() >= 1.0)
- def test_negative_value_z_operator(self):
- """
- Test the example given in Gowda/Ravindran of a Z-matrix with
- negative game value on the nonnegative orthant.
- """
- K = NonnegativeOrthant(2)
- e1 = [1, 1]
- e2 = e1
- L = [[1, -2], [-2, 1]]
- G = SymmetricLinearGame(L, K, e1, e2)
- self.assertTrue(G.solution().game_value() < -options.ABS_TOL)
-
-
- def assert_scaling_works(self, game1):
+ def assert_scaling_works(self, G):
"""
Test that scaling ``L`` by a nonnegative number scales the value
of the game by the same number.
"""
- value1 = game1.solution().game_value()
-
- alpha = uniform(0.1, 10)
- game2 = SymmetricLinearGame(alpha*game1._L.trans(),
- game1._K,
- game1._e1,
- game1._e2)
-
- while game2._condition() > MAX_COND:
- # Loop until the condition number of game2 doesn't suck.
- alpha = uniform(0.1, 10)
- game2 = SymmetricLinearGame(alpha*game1._L.trans(),
- game1._K,
- game1._e1,
- game1._e2)
-
- value2 = game2.solution().game_value()
- self.assert_within_tol(alpha*value1, value2)
+ (alpha, H) = random_nn_scaling(G)
+ value1 = G.solution().game_value()
+ value2 = H.solution().game_value()
+ modifier = 4*max(abs(alpha), 1)
+ self.assert_within_tol(alpha*value1, value2, modifier)
def test_scaling_orthant(self):
self.assert_scaling_works(G)
- def assert_translation_works(self, game1):
+ def assert_translation_works(self, G):
"""
Check that translating ``L`` by alpha*(e1*e2.trans()) increases
the value of the associated game by alpha.
"""
# We need to use ``L`` later, so make sure we transpose it
# before passing it in as a column-indexed matrix.
- soln1 = game1.solution()
+ soln1 = G.solution()
value1 = soln1.game_value()
x_bar = soln1.player1_optimal()
y_bar = soln1.player2_optimal()
- tensor_prod = game1._e1*game1._e2.trans()
# This is the "correct" representation of ``M``, but COLUMN
# indexed...
- alpha = uniform(-10, 10)
- M = game1._L + alpha*tensor_prod
-
- # so we have to transpose it when we feed it to the constructor.
- game2 = SymmetricLinearGame(M.trans(), game1._K, game1._e1, game1._e2)
- while game2._condition() > MAX_COND:
- # Loop until the condition number of game2 doesn't suck.
- alpha = uniform(-10, 10)
- M = game1._L + alpha*tensor_prod
- game2 = SymmetricLinearGame(M.trans(),
- game1._K,
- game1._e1,
- game1._e2)
-
- value2 = game2.solution().game_value()
+ (alpha, H) = random_translation(G)
+ value2 = H.solution().game_value()
- self.assert_within_tol(value1 + alpha, value2)
+ modifier = 4*max(abs(alpha), 1)
+ self.assert_within_tol(value1 + alpha, value2, modifier)
# Make sure the same optimal pair works.
- self.assert_within_tol(value2, inner_product(M*x_bar, y_bar))
+ self.assert_within_tol(value2, H.payoff(x_bar, y_bar), modifier)
def test_translation_orthant(self):
self.assert_translation_works(G)
- def assert_opposite_game_works(self, game1):
+ def assert_opposite_game_works(self, G):
"""
Check the value of the "opposite" game that gives rise to a
value that is the negation of the original game. Comes from
"""
# This is the "correct" representation of ``M``, but
# COLUMN indexed...
- M = -game1._L.trans()
+ M = -G.L().trans()
# so we have to transpose it when we feed it to the constructor.
- # Note: the condition number of game2 should be comparable to game1.
- game2 = SymmetricLinearGame(M.trans(), game1._K, game1._e2, game1._e1)
+ # Note: the condition number of ``H`` should be comparable to ``G``.
+ H = SymmetricLinearGame(M.trans(), G.K(), G.e2(), G.e1())
- soln1 = game1.solution()
+ soln1 = G.solution()
x_bar = soln1.player1_optimal()
y_bar = soln1.player2_optimal()
- soln2 = game2.solution()
+ soln2 = H.solution()
- self.assert_within_tol(-soln1.game_value(), soln2.game_value())
+ # The modifier of 4 is because each could be off by 2*ABS_TOL,
+ # which is how far apart the primal/dual objectives have been
+ # observed being.
+ self.assert_within_tol(-soln1.game_value(), soln2.game_value(), 4)
+
+ # Make sure the switched optimal pair works. Since x_bar and
+ # y_bar come from G, we use the same modifier.
+ self.assert_within_tol(soln2.game_value(), H.payoff(y_bar, x_bar), 4)
- # Make sure the switched optimal pair works.
- self.assert_within_tol(soln2.game_value(),
- inner_product(M*y_bar, x_bar))
def test_opposite_game_orthant(self):
y_bar = soln.player2_optimal()
value = soln.game_value()
- ip1 = inner_product(y_bar, G._L*x_bar - value*G._e1)
- self.assert_within_tol(ip1, 0)
+ ip1 = inner_product(y_bar, G.L()*x_bar - value*G.e1())
+ ip2 = inner_product(value*G.e2() - G.L().trans()*y_bar, x_bar)
- ip2 = inner_product(value*G._e2 - G._L.trans()*y_bar, x_bar)
- self.assert_within_tol(ip2, 0)
+ # Huh.. well, y_bar and x_bar can each be epsilon away, but
+ # x_bar is scaled by L, so that's (norm(L) + 1), and then
+ # value could be off by epsilon, so that's another norm(e1) or
+ # norm(e2). On the other hand, this test seems to pass most of
+ # the time even with a modifier of one. How about.. four?
+ self.assert_within_tol(ip1, 0, 4)
+ self.assert_within_tol(ip2, 0, 4)
def test_orthogonality_orthant(self):
This test theoretically applies to the ice-cream cone as well,
but we don't know how to make positive operators on that cone.
"""
- G = random_orthant_game()
- L = random_nonnegative_matrix(G._K.dimension())
-
- # Replace the totally-random ``L`` with the random nonnegative one.
- G = SymmetricLinearGame(L, G._K, G._e1, G._e2)
-
- while G._condition() > MAX_COND:
- # Try again until the condition number is satisfactory.
- G = random_orthant_game()
- L = random_nonnegative_matrix(G._K.dimension())
- G = SymmetricLinearGame(L, G._K, G._e1, G._e2)
-
+ G = random_positive_orthant_game()
self.assertTrue(G.solution().game_value() >= -options.ABS_TOL)
#
# See :meth:`assert_within_tol` for an explanation of the
# fudge factors.
- eigs = eigenvalues_re(G._L)
- epsilon = 2*2*RANDOM_MAX*options.ABS_TOL
- if soln.game_value() > epsilon:
+ eigs = eigenvalues_re(G.L())
+
+ if soln.game_value() > options.ABS_TOL:
# L should be positive stable
positive_stable = all([eig > -options.ABS_TOL for eig in eigs])
- if not positive_stable:
- print(str(eigs))
self.assertTrue(positive_stable)
- elif soln.game_value() < -epsilon:
+ elif soln.game_value() < -options.ABS_TOL:
# L should be negative stable
negative_stable = all([eig < options.ABS_TOL for eig in eigs])
- if not negative_stable:
- print(str(eigs))
self.assertTrue(negative_stable)
# The dual game's value should always equal the primal's.
+ # The modifier of 4 is because even though the games are dual,
+ # CVXOPT doesn't know that, and each could be off by 2*ABS_TOL.
dualsoln = G.dual().solution()
- self.assert_within_tol(dualsoln.game_value(), soln.game_value())
+ self.assert_within_tol(dualsoln.game_value(), soln.game_value(), 4)
def test_lyapunov_orthant(self):
"""
Test that a Lyapunov game on the nonnegative orthant works.
"""
- G = random_orthant_game()
- L = random_diagonal_matrix(G._K.dimension())
-
- # Replace the totally-random ``L`` with random Lyapunov-like one.
- G = SymmetricLinearGame(L, G._K, G._e1, G._e2)
-
- while G._condition() > MAX_COND:
- # Try again until the condition number is satisfactory.
- G = random_orthant_game()
- L = random_diagonal_matrix(G._K.dimension())
- G = SymmetricLinearGame(L, G._K, G._e1, G._e2)
-
+ G = random_ll_orthant_game()
self.assert_lyapunov_works(G)
"""
Test that a Lyapunov game on the ice-cream cone works.
"""
- G = random_icecream_game()
- L = random_lyapunov_like_icecream(G._K.dimension())
-
- # Replace the totally-random ``L`` with random Lyapunov-like one.
- G = SymmetricLinearGame(L, G._K, G._e1, G._e2)
-
- while G._condition() > MAX_COND:
- # Try again until the condition number is satisfactory.
- G = random_orthant_game()
- L = random_lyapunov_like_icecream(G._K.dimension())
- G = SymmetricLinearGame(L, G._K, G._e1, G._e2)
-
+ G = random_ll_icecream_game()
self.assert_lyapunov_works(G)