tf.multiply(tf.log(tf.gather_nd(l_out, actions_pl)), advantages_pl))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate_pl,
beta1=0.8,
beta2=0.92)
train_f = optimizer.minimize(loss_f)
saver = tf.train.Saver() # we use this later to save the model
# test forward pass
env = Minesweeper(display=False,
ROWS=6,
COLS=6,
MINES=7,
OUT="CONDENSED",
rewards={
"win": 1,
"loss": -1,
"progress": 0.9,
"noprogress": -0.3,
"YOLO": -0.3
})
state = env.stateConverter(env.get_state()).flatten()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
action_probabilities = sess.run(fetches=l_out,
feed_dict={states_pl: [state]})
print(action_probabilities)
# helper functions
rows = args.size
cols = args.size
mines = args.mine
FULL = False
rewards = {
"win": 0.9,
"loss": -1,
"progress": 0.9,
"noprogress": -0.3,
"YOLO": -0.3
}
env = Minesweeper(display=False,
FULL=FULL,
ROWS=rows,
COLS=cols,
MINES=mines,
rewards=rewards)
n_inputs = rows * cols * 10 if FULL else rows * cols * 2
n_hidden = [rows * cols * 10, 250, 250, 250, 250]
n_outputs = rows * cols
# Model
model = Sequential()
model.add(
Dense(
input_shape=(1, n_inputs),
units=n_hidden[0],
activation='relu',
kernel_initializer='glorot_uniform',
parser.add_argument(
"-d",
"--display",
help=
"run with this argument to display the game and see the agent play by pressing enter"
)
args = parser.parse_args()
if args.display:
display = True
env = Minesweeper(display=display,
ROWS=6,
COLS=6,
MINES=6,
OUT="FULL",
rewards={
"win": 1,
"loss": -1,
"progress": 0.9,
"noprogress": -0.3,
"YOLO": -0.4
})
with tf.Session() as sess:
#Restore Model
saver = tf.train.Saver()
saver.restore(sess, "{}/{}.ckpt".format(model, model))
#Initialize test parameters
games = 0
moves = 0
stuck = 0
