def build_graph():
# Create shared global policy and value networks
s, p_network, v_network, p_params, v_params = build_policy_and_value_networks(num_actions=ACTIONS, agent_history_length=AGENT_HISTORY_LENGTH, resized_width=RESIZED_WIDTH, resized_height=RESIZED_HEIGHT)
# Shared global optimizer
optimizer = tf.train.AdamOptimizer(LEARNING_RATE)
# Op for applying remote gradients
R_t = tf.placeholder("float", [None])
a_t = tf.placeholder("float", [None, ACTIONS])
log_prob = tf.log(tf.reduce_sum(tf.mul(p_network, a_t), reduction_indices=1))
p_loss = -log_prob * (R_t - v_network)
v_loss = tf.reduce_mean(tf.square(R_t - v_network))
total_loss = p_loss + (0.5 * v_loss)
minimize = optimizer.minimize(total_loss)
return s, a_t, R_t, minimize
def build_graph():
# Create shared global policy and value networks
(s, p_network, v_network, p_params,
v_params) = build_policy_and_value_networks(num_actions=NUM_ACTIONS,
input_shape=STATE_SHAPE)
# Shared global optimizer
optimizer = tf.train.RMSPropOptimizer(learning_rate=LEARNING_RATE,
decay=DECAY)
# Op for applying remote gradients
R_t = tf.placeholder("float", [None])
a_t = tf.placeholder("float", [None, NUM_ACTIONS])
log_prob = tf.log(tf.reduce_sum(p_network * a_t, axis=1))
p_loss = -log_prob * (R_t - v_network)
v_loss = tf.reduce_mean(tf.square(R_t - v_network))
total_loss = p_loss + (0.5 * v_loss)
minimize = optimizer.minimize(total_loss)
return s, a_t, R_t, minimize, p_network, v_network
