print("[MESSAGE] Model built.")
# training schedule
reward_sum = 0
running_reward = None
episode_number = 0
xs, dlogps, drs, probs = [], [], [], []
train_X, train_Y = [], []
num_victory = 0
# go through entire game space
while True:
for game_idx in xrange(num_train):
for start_pos in [start_tot[game_idx][0]]:
game = grid.Grid(data[game_idx],
value[game_idx],
imsize,
start_pos,
is_po=False)
# until the game is failed
while True:
# game_state = game.get_state()
# plt.subplot(1, 3, 1)
# plt.imshow(game_state[0, 0], cmap="gray")
# plt.subplot(1, 3, 2)
# plt.imshow(game_state[0, 1], cmap="gray")
# plt.subplot(1, 3, 3)
# plt.imshow(game_state[0, 2], cmap="gray")
# plt.show()
# print (game_state[0, 0])
# compute probability
aprob = model.predict(game.get_state()).flatten()
agent = DQN()
total_win = 0.
total_games = 0.
for game_idx in xrange(num_train):
for start_idx in range(len(start_tot[game_idx])):
print(start_tot[game_idx])
start_pos = start_tot[game_idx][start_idx]
curr_win = 0
start_pos_flag = True
for episode in xrange(EPISODE):
total_games += 1.
print ("\nThis is game %d, start position %d, %s" % (game_idx + 1, start_idx + 1, map(str,start_pos)))
print ("Start position is marked 33; Goal is marked 99; Other index indicate the latest step number; 1 is free space; 0 is obstacle.")
game = grid.Grid(data[game_idx], value[game_idx], imsize, start_pos = start_pos, is_po=True)
if start_pos_flag:
if start_pos != game.pos_history[0]:
start_pos = game.pos_history[0]
print (game.pos_history[0])
start_pos_flag = False
# Train
for step in xrange(STEP):
#update game state
if step == 0:
tmp_value_pos = value[game_idx].copy()
tmp_value_pos.reshape(dim, dim)[game.curr_pos] = 1
state = np.array([game.curr_map.ravel(), tmp_value_pos]).transpose()
else:
state = next_state
# get next action from current state
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
K.set_session(sess)
actor = ActorNet(sess, state_dim, action_dim, batch_size, tau, lra)
critic = CriticNet(sess, state_dim, action_dim, batch_size, tau, lrc)
buff = ReplayBuffer(buffer_size)
# the game loop
for game_idx in xrange(episode_count):
print("Episode : " + str(game_idx) + " Replay Buffer " + str(buff.count()))
for start_pos in [start_tot[0][0]]:
# start game
game = grid.Grid(data[0], value[0], imsize, start_pos, is_po=False)
done = False
s_t = game.get_state()
s_t = s_t.transpose((0, 2, 3, 1))
total_reward = 0.
while True:
# plt.subplot(1, 3, 1)
# plt.imshow(s_t[0, :, :, 0], cmap="gray")
# plt.subplot(1, 3, 2)
# plt.imshow(s_t[0, :, :, 1], cmap="gray")
# plt.show()
loss = 0
epsilon -= 1.0 / explore