def test_full2(self):
initial_sigmas = np.array([10, 12, 12, 12, 14], dtype=np.int32)
alpha = utils.borda(5) * 2 # kind-of-Borda
weights = np.array([2, 2])
m = len(initial_sigmas)
assert isinstance(initial_sigmas, np.ndarray)
gaps = utils.sigmas_to_gaps(initial_sigmas, np.max(initial_sigmas))
logger.info('gaps={}'.format(gaps))
init_gaps = gaps
t1 = current_milli_time()
fractional_makespan, clp_res = clp_R_alpha_WCM.find_strategy(
initial_sigmas, alpha, weights, mode='per_cand')
t2 = current_milli_time()
logger.warning('Took time {}ms, fractional_makespan={}'.format(
t2 - t1, fractional_makespan))
# fractional_makespan = utils.weighted_makespan(frac_res, alpha, weights, initial_sigmas)
clp_makespan = utils.weighted_makespan(clp_res, alpha, weights,
initial_sigmas)
logger.info('weights={} m={}frac={} CLP={}'.format(
weights, m, fractional_makespan, clp_makespan))
self.assertLessEqual(fractional_makespan, clp_makespan)
def find_strategy(initial_sigmas, alpha, weights):
if len(initial_sigmas) != len(alpha):
raise ValueError(len(initial_sigmas) != len(alpha))
m = len(initial_sigmas)
k = len(weights)
# order voter according to weights, non-increasing
ordered_voters = sorted(np.arange(k),
key=lambda v: weights[v],
reverse=True)
current_awarded = initial_sigmas.copy()
config_mat = np.zeros((m, k), dtype=int)
for ell in ordered_voters:
# now sort the candidates according to the current awarded score, high to low:
candidates_sorted = sorted(utils.borda(m),
key=lambda i: current_awarded[i],
reverse=True)
# now achieve inverse-sort: every candidate gets his rank (high-to-low) in sort
ballot = np.zeros(m, dtype=int)
for rank, cand in enumerate(candidates_sorted):
ballot[cand] = rank
config_mat[:, ell] = ballot
current_awarded += weights[ell] * alpha[ballot]
return config_mat
def test_calculate_awarded(self):
m = 3
config_mat = np.array([[0, 1], [2, 1], [2, 0]])
initil_sigmas = np.array([10, 11, 12])
weights = np.array([1, 2])
alpha = utils.borda(m) # borda
awarded = utils.weighted_calculate_awarded(config_mat, alpha, weights,
initil_sigmas)
self.assertEquals(awarded.tolist(), [12, 15, 14])
def test_find_strategy(self):
initial_sigmas = np.array([5, 6, 6, 6, 7], dtype=np.int32)
weights = np.array([1, 1])
k = len(weights)
m = len(initial_sigmas)
alpha = utils.borda(m)
res = reverse_algorithm.find_strategy(initial_sigmas, alpha, weights)
self.assertListEqual(np.ravel(res).tolist(), [4, 0, 1, 3, 2, 2, 3, 1, 0, 4])
def test_lp_solve(self):
initial_sigmas = np.array([4, 6, 6, 8], dtype=np.int32)
alpha = utils.borda(4) # Borda
weights = np.array([2, 2])
m = len(initial_sigmas)
target = 12
result = clp_R_alpha_WCM.lp_solve(m, alpha, weights, initial_sigmas,
target)
tol = 0.001
makespan = utils.weighted_fractional_makespan(initial_sigmas, result,
alpha, weights)
self.assertLessEqual(makespan, target + tol)
def test_fix_rounding_result_weighted(self):
k = 2
m = 3
config_mat = np.array([[0, 1], [2, 2], [2, 0]])
initial_sigmas = np.array([0, 1, 1])
weights = np.array([1, 1])
alpha = utils.borda(m) # borda
res_config_mat = clp_R_alpha_WCM.fix_rounding_result_weighted(
config_mat, alpha, weights, initial_sigmas)
self.assertListEqual(
np.ravel(res_config_mat).tolist(), [0, 1, 1, 2, 2, 0])
def experiments():
for m in range(10, 110, 10):
for k in [10]:
for trial in range(trials):
n = k * 2
ceil_log_n = int(np.ceil(np.log(n)))
yield delayed(run)(
n, utils.borda(m),
utils.draw_zipf_weights(owners=k,
W=k * ceil_log_n), m, trial,
utils.draw_uniform_weighted(
m, k * 2,
utils.draw_zipf_weights(owners=n,
W=n * ceil_log_n,
return_len_k=True)))
def test_full(self):
initial_sigmas = np.array([5, 6, 6, 6, 7], dtype=np.int32)
k = 2
m = len(initial_sigmas)
alpha = utils.borda(m) # borda
assert isinstance(initial_sigmas, np.ndarray)
gaps = utils.sigmas_to_gaps(initial_sigmas, np.max(initial_sigmas))
logger.info('gaps={}'.format(gaps))
init_gaps = gaps
t1 = current_milli_time()
fractional_makespan, clp_res = clp_R_alpha_UCM.find_strategy(
initial_sigmas, alpha, k, mode='per_cand')
t2 = current_milli_time()
logger.warning('Took time {}ms, fractional_makespan={}'.format(
t2 - t1, fractional_makespan))
# fractional_makespan = utils.makespan(initial_sigmas, frac_res, alpha=alpha)
clp_makespan = utils.makespan(initial_sigmas, clp_res, alpha=alpha)
gaps = utils.sigmas_to_gaps(initial_sigmas, clp_makespan)
af_res = average_fit.solve_one_gaps(k, gaps, verbose=False)
af_makespan = utils.makespan(initial_sigmas, af_res, alpha=alpha)
logger.info('k={} m={} frac={} CLP={} AF={}'.format(
k, m, fractional_makespan, clp_makespan, af_makespan))
# self.assertListEqual(clp_res.tolist(),
# [[0, 0, 1, 1, 0],
# [0, 1, 0, 1, 0],
# [1, 0, 0, 0, 1],
# [1, 0, 0, 0, 1],
# [0, 1, 1, 0, 0]]
# )
self.assertAlmostEqual(fractional_makespan, 10.0, delta=1e-3)
self.assertEquals(clp_makespan, 10)
self.assertEquals(af_makespan, 12)
def solve_one_gaps(k, init_gaps, verbose=False):
m = len(init_gaps)
score2mult = np.array([k] * m)
init_gaps = np.array(init_gaps, dtype=float)
gaps = init_gaps.copy()
left = np.array([k] * m)
potential = gaps / left
num_scores = m * k
config_mat = np.zeros((m, m), dtype=int)
if verbose:
strs = [
'{}/{}={}'.format(g, l, p)
for g, l, p in zip(gaps, left, potential)
]
print('\t\t'.join(strs))
while num_scores > 0:
# choose cand with highest potential
cand = np.argmax(potential)
# choose which score to give
does_fit = utils.borda(m) <= gaps[cand]
still_left = score2mult > 0
does_fit_and_still_left = np.logical_and(does_fit, still_left)
score_to_give = None
if np.any(does_fit_and_still_left):
score_to_give = does_fit_and_still_left.nonzero()[0][-1]
else:
score_to_give = still_left.nonzero()[0][-1]
# does_fit = np.array([ j <= gaps[cand] for j in range(m)])
# still_left = np.array([score2mult[j] > 0 for j in range(m)])
# if we have still score of this type to give, and it doesn't cause the candidate to overflow
score2mult[score_to_give] -= 1
config_mat[cand, score_to_give] += 1
gaps[cand] -= score_to_give
left[cand] -= 1
# re-calculate potential
if left[cand] > 0:
potential[cand] = gaps[cand] / left[cand]
else:
potential[cand] = -1
num_scores -= 1
if verbose:
strs = [
'{}/{}={}'.format(g, l, p)
for g, l, p in zip(gaps, left, potential)
]
strs[cand] = utils.bcolors.OKBLUE + str(
score_to_give
) + '->' + utils.bcolors.WARNING + strs[cand] + utils.bcolors.ENDC
print('\t\t'.join(strs))
logging.debug('final configs:')
logging.debug(config_mat)
return config_mat