def train():
env = my_env.MyEnv(0, realtime_mode=True)
model = CatBoostClassifier()
model.load_model("catboost_model.model")
score = 0.0
print_interval = 1
for n_epi in range(10000):
s = env.reset()
done = False
while not done:
y_pred1 = model.predict(s, prediction_type="Probability")
if deterministic:
y_pred_max = int(np.argmax(y_pred1))
a = action_mapping(y_pred_max)
else:
a = int(np.random.choice([0, 1, 3, 4, 5], p=y_pred1))
s_prime, r, done, info = env.step(a)
s = s_prime
score += r
if done:
break
if n_epi%print_interval==0 and n_epi!=0:
print("# of episode :{}, avg score : {:.5f}".format(n_epi, score/print_interval))
score = 0.0
env.close()
def train():
env = my_env.MyEnv(0, realtime_mode=True)
model = bc.BC()
model.load_state_dict(torch.load("imitation_model_1000.pt"))
score = 0.0
print_interval = 1
for n_epi in range(10000):
s = env.reset()
done = False
while not done:
prob = model(torch.from_numpy(s).float())
m = Categorical(prob)
a = m.sample().item()
s_prime, r, done, info = env.step(a)
s = s_prime
score += r
if done:
break
if n_epi % print_interval == 0 and n_epi != 0:
print("# of episode :{}, avg score : {:.5f}".format(
n_epi, score / print_interval))
score = 0.0
env.close()
def train():
env = my_env.MyEnv(0)
model = PPO()
score = 0.0
print_interval = 1
for n_epi in range(10000):
s = env.reset()
done = False
while not done:
for t in range(T_horizon):
a_prob, pi_b1, pi_b2, pi_b3 = model.pi(
torch.from_numpy(s).float())
a_m = Categorical(a_prob)
a = a_m.sample().item()
if a == 0:
b1_m = Categorical(pi_b1)
b1 = b1_m.sample().item()
b = b1
a_b_prob = a_prob[a] * pi_b1[b1]
elif a == 1:
b2_m = Categorical(pi_b2)
b2 = b2_m.sample().item()
b = b2
a_b_prob = a_prob[a] * pi_b2[b2]
else:
b3_m = Categorical(pi_b3)
b3 = b3_m.sample().item()
b = b3
a_b_prob = a_prob[a] * pi_b3[b3]
action = a * 3 + b
s_prime, r, done, info = env.step(action)
model.put_data(
(s, a, b, r, s_prime, a_b_prob, np.eye(3)[a], done))
s = s_prime
score += r
if done:
break
model.train_net()
if n_epi % 50 == 0 and n_epi != 0:
torch.save(model.state_dict(), f"ppo_model_{n_epi}.pt")
print("saved!")
break
if n_epi % print_interval == 0 and n_epi != 0:
print("# of episode :{}, avg score : {:.1f}".format(
n_epi, score / print_interval))
score = 0.0
env.close()
def train():
env = my_env.MyEnv(0)
model1 = PPO()
model2 = PPO()
score = 0.0
print_interval = 20
for n_epi in range(10000):
s = env.reset()
done = False
while not done:
for t in range(T_horizon):
prob1 = model1.pi(torch.from_numpy(s).float())
prob2 = model2.pi(torch.from_numpy(s).float())
m1 = Categorical(prob1)
m2 = Categorical(prob2)
a1 = m1.sample().item()
a2 = m2.sample().item()
s_prime, r, done, info = env.step(np.array([a1, a2]))
model1.put_data((s, a1, r, s_prime, prob1[a1].item(), done))
model2.put_data((s, a2, r, s_prime, prob2[a2].item(), done))
s = s_prime
score += r
if done:
break
model1.train_net()
model2.train_net()
# if n_epi % 50 == 0 and n_epi != 0:
# torch.save(model.state_dict(), f"ppo_model_{n_epi}.pt")
# print("saved!")
# break
if n_epi % print_interval == 0 and n_epi != 0:
print("# of episode :{}, avg score : {:.1f}".format(
n_epi, score / print_interval))
score = 0.0
env.close()
def train():
env = my_env.MyEnv(0)
model = PPO()
score = 0.0
print_interval = 20
for n_epi in range(10000):
s = env.reset()
done = False
while not done:
for t in range(T_horizon):
prob = model.pi(torch.from_numpy(s).float())
m = Categorical(prob)
a = m.sample().item()
s_prime, r, done, info = env.step(a)
model.put_data((s, a, r, s_prime, prob[a].item(), done))
s = s_prime
score += r
if done:
break
model.train_net()
if n_epi % 50 == 0 and n_epi != 0:
torch.save(model.state_dict(), f"ppo_model_{n_epi}.pt")
print("saved!")
break
if n_epi % print_interval == 0 and n_epi != 0:
print("# of episode :{}, avg score : {:.1f}".format(
n_epi, score / print_interval))
score = 0.0
env.close()
print('LOSE')
reward = FAIL_PENALTY
agent.learn(cur_state, action, next_state, reward, done)
print("Episode finished after {} timesteps".format(t + 1))
print('#################################\n')
history.append(t + 1)
break
agent.learn(cur_state, action, next_state, reward, done)
cur_state = next_state
if t == MAX_STEPS - 1:
history.append(t + 1)
print("Episode finished after {} timesteps".format(t + 1))
return agent, history
env = my_env.MyEnv() #gym.make('CartPole-v0') # TODO myenv
# if RECORD:
# env = wrappers.Monitor(env, '/home/vbalogh/git/reinforcement_learning-stormmax/my-experiment-1', force=True)
def get_actions(current_state):
all_actions = list(range(my_env.N_EDGES))
if 0 in current_state:
current_pos = current_state.index(0)
del all_actions[current_pos]
return all_actions
agent = qlearning.QLearningAgent(get_actions,
epsilon=EPSILON,
alpha=LEARNING_RATE,