def test_add_subtract(self):
d1 = strategy.MoveCountDistribution([0.75, 0.25])
d2 = strategy.MoveCountDistribution([0.7, 0.2, 0.1])
np.testing.assert_allclose((d1 + d2).dist, [1.45, 0.45, 0.1])
np.testing.assert_allclose((d2 + d1).dist, [1.45, 0.45, 0.1])
np.testing.assert_allclose((d1 - d2).dist, [0.05, 0.05, -0.1])
np.testing.assert_allclose((d2 - d1).dist, [-0.05, -0.05, 0.1])
def test_expected_value(self):
np.testing.assert_approx_equal(
strategy.MoveCountDistribution([0.75, 0.25]).expected_value(),
0.25)
np.testing.assert_approx_equal(
strategy.MoveCountDistribution([0.0, 0.2, 0.8]).expected_value(),
1.8)
np.testing.assert_approx_equal(
strategy.MoveCountDistribution([0.0, 0.0, 1.0]).expected_value(),
2)
def test_trim_low_prob(self):
np.testing.assert_allclose(
strategy.MoveCountDistribution([0.1, 0.2,
0.3]).trim_low_prob(0.01).dist,
[0.1, 0.2, 0.3])
np.testing.assert_allclose(
strategy.MoveCountDistribution([0.1, 0.2,
0.3]).trim_low_prob(0.15).dist,
[0.0, 0.2, 0.3])
np.testing.assert_allclose(
strategy.MoveCountDistribution([0.3, 0.2, 0.1,
0.01]).trim_low_prob(0.15).dist,
[0.3, 0.2])
def create_distribution_store_from_gnubg(gnubg_dir):
"""Creates a database in our format from the gnubg database.
This uses a really dumb and inefficient strategy of calling
'bearoffdump' many times as separate processes. But it allows us
to not tweak the gnubg code and not have to deal with the vagaries
of their format (just some annoying text parsing)
Args:
gnubg_dir: source directory of gnubg with everyting compiled
Returns:
strategy.DistributionStore
"""
progress_interval = 500
config = board.GameConfiguration(15, 6)
store = strategy.DistributionStore(config)
start_time = time.time()
# gnubg uses a 1 based index as an argument to
# bearoffdump. However, it doesn't have the end state as a valid
# index so we add that manually.
store.distribution_map[config.min_board_id] = (
strategy.MoveCountDistribution([1]))
for idx in range(1, config.num_valid_boards):
completed_process = subprocess.run([
os.path.join(gnubg_dir, 'bearoffdump'),
os.path.join(gnubg_dir, 'gnubg_os0.bd'), '-n',
str(idx)
],
universal_newlines=True,
stdout=subprocess.PIPE,
check=True)
try:
b, mcd = parse_gnubg_dump(config, completed_process.stdout)
except ValueError as err:
raise ValueError('For gnubg index {}: {}'.format(idx, err))
store.distribution_map[b.get_id()] = mcd
if idx % progress_interval == 0:
frac_complete = idx / config.num_valid_boards
this_time = time.time()
print("%d/%d %.1f%%, %fs elapsed, %fs estimated total" %
(idx, config.num_valid_boards, frac_complete * 100,
this_time - start_time,
(this_time - start_time) / frac_complete),
flush=True)
return store
def parse_gnubg_dump(config, gnubg_str):
"""Parse the output of gnubg's "bearoffdump".
This parses the "Opponent" move distribution from thet "Bearing off" column
Args:
config: board.GameConfiguration
gnubg_str: string to parse
Returns
board.Board, strategy.MoveCountDistribution
"""
pos_id_str = None
mcd = None
parsing_mcd = False
pos_id_re = re.compile(r'GNU Backgammon Position ID: ([A-Za-z0-9+/]*)')
start_mcd_re = re.compile(r'^Rolls\s+Player\s+Opponent')
mcd_line_re = re.compile(r'^\s*\d+\s+[\d\.]+\s+([\d\.]+)')
for line in gnubg_str.splitlines():
if parsing_mcd:
match = mcd_line_re.search(line)
if match:
value = float(match.group(1)) / 100.0
if not mcd:
mcd = strategy.MoveCountDistribution([value])
else:
mcd = mcd.append([value])
else:
parsing_mcd = False
elif start_mcd_re.search(line):
parsing_mcd = True
else:
match = pos_id_re.search(line)
if match:
pos_id_str = match.group(1)
if not pos_id_str:
raise ValueError('Never found position id line')
if not mcd:
raise ValueError('Never found move distribution')
return (board.Board.from_id(config,
gnubg_id_str_to_board_id(config,
pos_id_str)), mcd)
def testParse(self):
input = """
Bearoff database: gnubg_os0.bd
Position number : 30
Information about database:
* On disk 1-sided bearoff database evaluator
- generated by GNU Backgammon
- up to 15 chequers on 6 points (54264 positions) per player
- database includes gammon distributions
Dump of position#: 30
GNU Backgammon Position ID: EwAAAAAAAAAAAA
+13-14-15-16-17-18------19-20-21-22-23-24-+
| | | O O | OOO
| | | O | OOO
| | | | OO
| | | | OO
| | | | OO
v| |BAR| |
| | | | XXX
| | | | XXX
| | | | XXX
| | | | XXX
| | | | XXX
+12-11-10--9--8--7-------6--5--4--3--2--1-+
Player Opponent
Position 0 30
Bearing off Bearing at least one chequer off
Rolls Player Opponent Player Opponent
0 100.000 0.000 100.000 100.000
1 0.000 13.889 0.000 0.000
2 0.000 86.111
Average rolls
Bearing off Saving gammon
Player Opponent Player Opponent
Mean 0.000 1.861 0.000 0.000
Std dev 0.000 0.346 0.000 0.000
Effective pip count:
Player Opponent
EPC 0.000 15.199
Wastage 0.000 10.199
EPC = 8.167 * Average rolls
Wastage = EPC - pips
"""
config = board.GameConfiguration(15, 6)
b, mcd = gnubg_interface.parse_gnubg_dump(config, input)
expected_b = board.Board(config, [12, 2, 0, 1, 0, 0, 0])
self.assertEqual(b,
expected_b,
msg='expected={}, got={}'.format(expected_b, b))
expected_mcd = strategy.MoveCountDistribution([0.0, .13889, .86111])
np.testing.assert_allclose(expected_mcd.dist, mcd.dist, atol=1e-5)
def test_append(self):
d = strategy.MoveCountDistribution([0.1, 0.2])
np.testing.assert_allclose(
d.append([0.3, 0.4]).dist, [0.1, 0.2, 0.3, 0.4])
def test_is_normalized(self):
self.assertTrue(
strategy.MoveCountDistribution([0.75, 0.25]).is_normalized())
self.assertFalse(
strategy.MoveCountDistribution([0.1, 0.2]).is_normalized())
def test_increase_counts(self):
d = strategy.MoveCountDistribution([0.75, 0.25])
np.testing.assert_allclose(
d.increase_counts(2).dist, [0, 0, 0.75, 0.25])
def test_multiply_divide(self):
d = strategy.MoveCountDistribution([0.75, 0.25])
np.testing.assert_allclose((d * 2).dist, [1.5, .5])
np.testing.assert_allclose((d / 5).dist, [.15, .05])
def test_init(self):
d = strategy.MoveCountDistribution()
np.testing.assert_allclose(d.dist, [0])
with self.assertRaises(ValueError):
strategy.MoveCountDistribution([[0.1], [0.2]])