def train_dqn():
args = DQNArgs()
env = gym.make(args.env_name)
agent = DQNAgent(env, QNet, SimpleNormalizer, args)
pre_best = -1e9
for ep in range(args.max_ep):
agent.train_one_episode()
if ep % args.test_interval == 0:
r = agent.test_model()
if r > pre_best:
pre_best = r
agent.save(args.save_dir)
if len(agent.memory) >= batch_size:
loss_sim_list.append(agent.replay(batch_size))
# For data visualisation
i.append(mdp.s[0, -1])
v.append(mdp.s[1, -1])
r.append(mdp.reward)
loss_over_simulation_time = np.sum(np.array([loss_sim_list])[0]) / len(
np.array([loss_sim_list])[0])
loss_of_episode.append(loss_over_simulation_time)
print("Initial Heading : {}".format(hdg0_rand))
print("----------------------------")
print("episode: {}/{}, Mean Loss = {}".format(e, EPISODES,
loss_over_simulation_time))
print("----------------------------")
agent.save("../Networks/dqn-test")
# plt.semilogy(np.linspace(1, EPISODES, EPISODES), np.array(loss_of_episode))
# plt.xlabel("Episodes")
# plt.ylabel("Cost")
f, axarr = plt.subplots(4, sharex=True)
axarr[0].plot(np.array(i[floor(len(i) / 2):len(i) - 1]) / TORAD)
axarr[1].plot(v[floor(len(i) / 2):len(i) - 1])
axarr[2].plot(r[floor(len(i) / 2):len(i) - 1])
axarr[3].semilogy(loss_sim_list[floor(len(i) / 2):len(i) - 1])
axarr[0].set_ylabel("angle of attack")
axarr[1].set_ylabel("v")
axarr[2].set_ylabel("r")
axarr[3].set_ylabel("cost")
plt.show()
state_size = 3
action_size = 9
actions = [ [[0,0],[-100,-100]], [[0,0],[-100,0]], [[0,0],[-100,100]],
[[0,0],[0,-100]], [[0,0],[0,0]], [[0,0],[0,100]],
[[0,0],[100,-100]], [[0,0],[100,0]], [[0,0],[100,100]]]
env = MyEnvironment()
agent = DQNAgent(state_size, action_size)
agent.load("./save/example_dqn.h5")#load
batch_size = 32
for e in range(3000):
state = np.reshape(env.reset(), [1, state_size])
last_reward = 0
for time in range(1000):
env.render() #render
action = agent.act_2(state)
commands = actions[action]
next_state, reward, done, _ = env.step2(commands)
next_state = np.reshape(next_state, [1, state_size])
agent.remember(state, action, reward, next_state, done)
state = next_state
if len(agent.memory) > batch_size:
agent.replay(batch_size)
if e % 1 == 0:
agent.save("./save/example_dqn.h5")
action = agent.act(state)
next_state, reward, done, _ = env.step(action)
# reward = reward if not done else -10
next_state = np.reshape(next_state, [1, state_size])
agent.remember(state, action, reward, next_state, done)
state = next_state
# print(action, reward)
if done:
print("episode: {}/{}, score: {}, e: {:.5}".format(
e, EPISODES, time, agent.epsilon))
break
if len(agent.memory) > batch_size:
agent.replay(batch_size)
if e % 10 == 0:
save_string = './save/' + stock_name + '_weights_with_fees.h5'
agent.save(save_string)
# # serialize model to JSON
# model_json = model.to_json()
# with open("model.json", "w") as json_file:
# json_file.write(model_json)
# # serialize weights to HDF5
# model.save_weights("model.h5")
# print("Saved model to disk")
# # later...
# # load json and create model
# json_file = open('model.json', 'r')
# loaded_model_json = json_file.read()
# json_file.close()
next_state = np.reshape(next_state, [1, state_size])
state = next_state
if (done):
reward = 2000
if (reward <= last_reward and done == False):
last_reward = reward
reward = -1000
else:
last_reward = reward
if (TRAINING):
agent.remember(state, action, reward, next_state, done)
if done:
print("1;{};{};{:.2f};{:.2}".format(e, EPISODES, reward,
agent.epsilon))
if (TRAINING):
agent.replay(len(agent.memory))
break
if len(agent.memory) > batch_size and TRAINING:
agent.replay(batch_size)
if (time == MOVES - 1):
print("0;{};{};{:.2f};{:.2}".format(e, EPISODES, last_reward,
agent.epsilon))
if (TRAINING):
agent.save("./save/execution1.h5")
