r_ = -1.0
replay_buffer.add(new_inputs, a, r_, new_inputs_)
else:
g_ = episode_experience[bit_size - 1][3]
new_inputs = np.concatenate([s, g_], axis=-1)
new_inputs_ = np.concatenate([s_, g_], axis=-1)
if (np.array(s_) == np.array(g_)).all():
r_ = 0.0
else:
r_ = -1.0
replay_buffer.add(new_inputs, a, r_, new_inputs_)
losses = []
for k in range(optimisation_steps):
state, action, reward, next_state = replay_buffer.sample()
target_net_Q = sess.run(target_model.out,
feed_dict={target_model.inp: next_state})
if FLAGS.DDQN:
main_net_predict = sess.run(model.predict,
feed_dict={model.inp: next_state})
doubleQ = np.reshape(
target_net_Q[range(main_net_predict.shape[0]),
main_net_predict], [-1])
target_reward = np.clip(
np.reshape(reward, [-1]) + gamma * doubleQ,
-1. / (1 - gamma), 0)
else:
target_reward = np.clip(
np.reshape(reward, [-1]) +
