Pownie/backgammon
1
0
Fork 0
Code Issues 6 Pull Requests Packages Projects Releases Wiki Activity

tesauro fat and diffs in values

Browse Source
This commit is contained in:
Alexander Munch-Hansen 2018-05-22 15:39:14 +02:00
parent 6e061171da
commit f170bad9b1
2 changed files with 6 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
board.py
View File

@ -268,23 +268,13 @@ class Board:
# print("Dice permuts:",dice_permutations) # print("Dice permuts:",dice_permutations)
for roll in dice_permutations: for roll in dice_permutations:
# Calculate boards resulting from first move # Calculate boards resulting from first move
#print("initial board: ", board)
#print("roll:", roll)
#print("Rest of roll:",roll[1:])
boards = calc_moves(board, roll[0]) boards = calc_moves(board, roll[0])
#print("Boards:",boards)
#print("Roll:",roll[0])
#print("boards after first die: ", boards)
for die in roll[1:]: for die in roll[1:]:
# Calculate boards resulting from second move # Calculate boards resulting from second move
nested_boards = [calc_moves(board, die) for board in boards] nested_boards = [calc_moves(board, die) for board in boards]
#print("nested boards: ", nested_boards)
boards = [board for boards in nested_boards for board in boards] boards = [board for boards in nested_boards for board in boards]
# What the fuck
#for board in boards:
# print(board)
# print("type__:",type(board))
# Add resulting unique boards to set of legal boards resulting from roll # Add resulting unique boards to set of legal boards resulting from roll
#print("printing boards from calculate_legal_states: ", boards) #print("printing boards from calculate_legal_states: ", boards)

6
network.py
View File

@ -114,12 +114,11 @@ class Network:
with tf.GradientTape() as tape: with tf.GradientTape() as tape:
value = self.model(prev_state.reshape(1,-1)) value = self.model(prev_state.reshape(1,-1))
grads = tape.gradient(value, self.model.variables) grads = tape.gradient(value, self.model.variables)
difference_in_values = tf.reshape(tf.subtract(value_next, value, name='difference_in_values'), []) difference_in_values = tf.reshape(tf.subtract(value_next, value, name='difference_in_values'), [])
tf.summary.scalar("difference_in_values", tf.abs(difference_in_values))
with tf.variable_scope('apply_gradients'):
for grad, train_var in zip(grads, self.model.variables): for grad, train_var in zip(grads, self.model.variables):
backprop_calc = self.learning_rate * difference_in_values * grad backprop_calc = self.learning_rate * difference_in_values * grad
train_var.assign_add(backprop_calc) train_var.assign_add(backprop_calc)
@ -299,7 +298,7 @@ class Network:
length_list = [] length_list = []
test_list = [] test_list = []
# Prepping of data # Prepping of data
start = time.time() # start = time.time()
for board in boards: for board in boards:
length = 0 length = 0
for roll in all_rolls: for roll in all_rolls:
@ -478,7 +477,6 @@ class Network:
for episode in range(1, episodes + 1): for episode in range(1, episodes + 1):
sys.stderr.write("[TRAIN] Episode {}".format(episode + trained_eps)) sys.stderr.write("[TRAIN] Episode {}".format(episode + trained_eps))
# TODO decide which player should be here
# player = 1 # player = 1
player = random.choice([-1,1]) player = random.choice([-1,1])