def sarsa_nstep_diff_live(W, Nruns):
print("Running nstep SARSA live")
for run in range(Nruns):
print("Run " + str(run + 1))
sim_environment.start_new_run(run)
curr_s = initial_state_generate()
t = 0
while True:
intersection = t % 4
if intersection == 3:
a_space = [13, 14, 15]
else:
a_space = [
4 * intersection + 1, 4 * intersection + 2,
4 * intersection + 3, 4 * intersection + 4
]
a = epsilon_greedy_a(0, a_space, curr_s, W)
env_param = sim_environment.take_action(a)
next_s = env_param['next_state']
r = env_param['rwd']
if r == 1000:
print("End of simulation at t = " + str(t))
break
curr_s = next_s
t += 1
return
def static_signalling(Nruns):
print("Running Static signalling")
for run in range(Nruns):
print("Run " + str(run + 1))
sim_environment.start_new_run(run)
initial_state_generate()
curr_a = 1 # cyclic test
t = 0
counter = [1, 0, 0, 0]
while True:
next_intersection = (t + 1) % 4
if next_intersection == 3:
a_space = [13, 14, 15]
else:
a_space = [
4 * next_intersection + 1, 4 * next_intersection + 2,
4 * next_intersection + 3, 4 * next_intersection + 4
]
next_a = a_space[counter[next_intersection]]
if counter[next_intersection] != len(a_space) - 1:
counter[next_intersection] += 1
else:
counter[next_intersection] = 0
env_param = sim_environment.take_action(curr_a)
r = env_param['rwd']
if r == 1000:
print("End of simulation at t = " + str(t))
break
curr_a = next_a
t += 1
return
def qr_dqn_live(load):
print("Running QR-DQN Live")
# delete any existing images
if platform.system() == 'Windows':
os.system("del .\img\*.png")
elif platform.system() == 'Linux':
os.system("rm .\img\*.png")
if load:
model = Network(len_state=STATE_LEN, num_quant=NUM_QUANTS, num_actions=NUM_ACTIONS)
model.load_state_dict(torch.load(TMPATH))
model.eval()
else:
model = Network(len_state=STATE_LEN, num_quant=NUM_QUANTS, num_actions=NUM_ACTIONS)
model.load_state_dict(Z.state_dict())
model.eval()
t = 0
sim_environment.start_new_run(0)
state = initial_state_generate()
plt.show()
plt.ion()
while True:
plt.clf()
plt.title('step = %s' % t)
intersection = t % 4
if intersection == 3:
a_space = [12, 13, 14]
else:
a_space = [4 * intersection, 4 * intersection + 1, 4 * intersection + 2,
4 * intersection + 3]
action = model.select_action(torch.Tensor([state]), a_space, 0)
observ = sim_environment.take_action(action + 1)
state = observ['next_state']
reward = observ['rwd']
done = 1 if reward == -100 else 0
t += 1
Zval = model(torch.Tensor([state])).detach().numpy()
for i in range(NUM_ACTIONS):
x, y = get_plot(Zval[0][i])
plt.plot(x, y, label='%s Q=%.1f' % (i + 1, Zval[0][i].mean()))
plt.legend(bbox_to_anchor=(1.1, 1.1), ncol=NUM_ACTIONS, prop={'size': 3})
if done: break
plt.savefig('./img/%s.png' % t)
plt.pause(0.001)
plt.close()
print("Steps = ", t)
return
def qr_dqn_live_noplots(Nruns, load):
print("Running QR-DQN Live (no plots)")
if load:
model = Network(len_state=STATE_LEN, num_quant=NUM_QUANTS, num_actions=NUM_ACTIONS)
model.load_state_dict(torch.load(TMPATH))
model.eval()
else:
model = Network(len_state=STATE_LEN, num_quant=NUM_QUANTS, num_actions=NUM_ACTIONS)
model.load_state_dict(Z.state_dict())
model.eval()
for run in range(Nruns):
t = 0
sim_environment.start_new_run(run)
state = initial_state_generate()
while True:
intersection = t % 4
if intersection == 3:
a_space = [12, 13, 14]
else:
a_space = [4 * intersection, 4 * intersection + 1, 4 * intersection + 2,
4 * intersection + 3]
action = model.select_action(torch.Tensor([state]), a_space, 0)
observ = sim_environment.take_action(action + 1)
state = observ['next_state']
reward = observ['rwd']
done = 1 if reward == -100 else 0
t += 1
if done: break
print("Steps = ", t)
return
def lqf(Nruns):
print("Running LQF")
for run in range(Nruns):
print("Run " + str(run + 1))
sim_environment.start_new_run(run)
initial_state_generate()
curr_a = random.randint(1, 4)
t = 0
while True:
next_intersection = (t + 1) % 4
env_param = sim_environment.take_action(curr_a)
next_s = env_param['next_state']
r = env_param['rwd']
if r == -100:
print("End of simulation at t = " + str(t))
break
if next_intersection == 3:
a_space = [13, 14, 15]
else:
a_space = [
4 * next_intersection + 1, 4 * next_intersection + 2,
4 * next_intersection + 3, 4 * next_intersection + 4
]
q_next = []
for a_temp in a_space:
q_next.append(next_s[a_temp - 1])
q_max = max(q_next)
q_max_index = [i for i, j in enumerate(q_next) if j == q_max]
rand_greedy_q = np.random.choice(q_max_index)
next_a = a_space[rand_greedy_q]
curr_a = next_a
t += 1
return
def qr_dqn_train(Nruns):
print("Running QR-DQN Training")
# Quantiles
tau = torch.Tensor((2 * np.arange(NUM_QUANTS) + 1) / (2.0 * NUM_QUANTS)).view(1, -1)
logger = Logger('q-net', fmt={'loss': '.5f'})
steps_done = 0
running_reward = 0
for run in range(Nruns):
t = 0
sum_reward = 0.0
memory = ReplayMemory(REPLAY_MEM_SIZE) # Initialize Replay buffer
sim_environment.start_new_run(run)
state = initial_state_generate()
while True:
intersection = t % 4
if intersection == 3:
a_space = [12, 13, 14]
else:
a_space = [4 * intersection, 4 * intersection + 1, 4 * intersection + 2,
4 * intersection + 3]
action = Z.select_action(torch.Tensor([state]), a_space, calc_epsilon(steps_done))
observ = sim_environment.take_action(action + 1)
next_state = observ['next_state']
reward = observ['rwd']
done = 1 if reward == -100 else 0
steps_done += 1
t += 1
if not done:
memory.push(state, action, next_state, reward, float(done))
sum_reward += reward
if len(memory) < BATCH_SIZE:
state = next_state
continue
states, actions, rewards, next_states, dones = memory.sample(BATCH_SIZE)
theta = Z(states)[np.arange(BATCH_SIZE), actions]
Znext = Ztgt(next_states).detach()
Qnext_sa = Znext.mean(2)
anext_max = torch.zeros([BATCH_SIZE], dtype=torch.long)
for i in range(BATCH_SIZE):
next_aspace = calc_next_aspace(int(actions[i]))
temp = Qnext_sa[i, :]
anext_max[i] = temp[next_aspace].max(0)[1] + next_aspace[0]
Znext_max = Znext[np.arange(BATCH_SIZE), anext_max]
Ttheta = rewards + GAMMA * (1 - dones) * Znext_max
diff = Ttheta.t().unsqueeze(-1) - theta
loss = huber(diff) * (tau - (diff.detach() < 0).float()).abs()
loss = loss.mean()
optimizer.zero_grad()
loss.backward()
optimizer.step()
state = next_state
if steps_done % NN_SYNC_FREQ == 0:
Ztgt.load_state_dict(Z.state_dict())
if done:
running_reward = sum_reward if not running_reward else 0.2*sum_reward + 0.8*running_reward
logger.add(run + 1, steps=t, running_reward=running_reward, loss=loss.data.numpy())
logger.iter_info()
break
torch.save(Z.state_dict(), TMPATH)
return
def sarsa_nstep_diff_train(n, c, epsilon, Nruns):
print("Running nstep SARSA training")
buff_len = n + 1
weight = np.zeros([S_LEN * A_LEN + 1, 1])
avg_reward = 0
dl_counter = 0
for run in range(Nruns):
print("Run " + str(run + 1))
r_arr = np.zeros(n, dtype=int)
a_arr = np.zeros(buff_len, dtype=int)
s_arr = []
alpha = 0.1 / (run + 1)
beta = c * alpha
e = math.exp(-run)
#Start new run
sim_environment.start_new_run(run)
curr_s = initial_state_generate()
curr_a = random.randint(1, 4)
a_arr[0] = curr_a
s_arr.insert(0, curr_s)
t = 0
while True:
next_intersection = (t + 1) % 4
#signals corresponding to that intersection
if next_intersection == 3:
a_space = [13, 14, 15]
else:
a_space = [
4 * next_intersection + 1, 4 * next_intersection + 2,
4 * next_intersection + 3, 4 * next_intersection + 4
]
#take action on intersection
env_param = sim_environment.take_action(curr_a)
#if all the traffic has left the simulation, sim_environment.py returns 1000 reward
r = env_param['rwd']
if r == 1000:
print("End of simulation at t = " + str(t))
break
next_s = env_param['next_state']
r_arr[t % n] = r
#chose next action epsilon greedily
next_a = epsilon_greedy_a(e, a_space, next_s, weight[:, 0])
#store the next state and next action
s_arr.insert((t + 1) % (n + 1), next_s)
a_arr[(t + 1) % (n + 1)] = next_a
tau = t - n + 1
#n-step SARSA Algo from Richard S. Sutton's Reinforcement Learning book
if tau >= 0:
q_tau_n = q_est(s_arr[(tau + n) % (n + 1)],
a_arr[(tau + n) % (n + 1)], weight[:, 0])
q_tau = q_est(s_arr[tau % (n + 1)], a_arr[tau % (n + 1)],
weight[:, 0])
do_error = sum(r_arr) - n * avg_reward + q_tau_n - q_tau
avg_reward = avg_reward + beta * do_error
phi_s_a_tau = phi(s_arr[tau % (n + 1)], a_arr[tau % (n + 1)])
weight[:, 0] = weight[:, 0] + alpha * do_error * np.transpose(
phi_s_a_tau)
curr_a = next_a
t += 1
W = weight[:, 0]
return W
def sarsa_nstep_diff_train(n, c, epsilon, Nruns):
print("Running nstep SARSA training")
buff_len = n + 1
r_arr = np.zeros(n, dtype=int)
a_arr = np.zeros(buff_len, dtype=int)
s_arr = []
weight = np.zeros([S_LEN * A_LEN + 1, 1])
for run in range(Nruns):
print("Run " + str(run + 1))
avg_reward = 10 # initialize avg reward
e = epsilon - run * (epsilon / Nruns)
if e < 0.4:
e = 0.4
sim_environment.start_new_run(run)
curr_s = initial_state_generate()
curr_a = random.randint(1, 4)
a_arr[0] = curr_a
s_arr.insert(0, curr_s)
t = 0
while True:
next_intersection = (t + 1) % 4
if next_intersection == 3:
a_space = [13, 14, 15]
else:
a_space = [
4 * next_intersection + 1, 4 * next_intersection + 2,
4 * next_intersection + 3, 4 * next_intersection + 4
]
alpha = 1 / (math.ceil((t + 1) / 10))
beta = c * alpha
env_param = sim_environment.take_action(curr_a)
r = env_param['rwd']
if r == -100:
print("Simulation time", t) # for test
break
next_s = env_param['next_state']
r_arr[t % n] = r
next_a = epsilon_greedy_a(e, a_space, next_s, weight[:, 0])
s_arr.insert((t + 1) % (n + 1), next_s)
a_arr[(t + 1) % (n + 1)] = next_a
tau = t - n + 1
if tau >= 0:
q_tau_n = q_est(s_arr[(tau + n) % (n + 1)],
a_arr[(tau + n) % (n + 1)], weight[:, 0])
q_tau = q_est(s_arr[tau % (n + 1)], a_arr[tau % (n + 1)],
weight[:, 0])
do_error = sum(r_arr) - n * avg_reward + q_tau_n - q_tau
avg_reward = avg_reward + beta * do_error
phi_s_a_tau = phi(s_arr[tau % (n + 1)], a_arr[tau % (n + 1)])
weight[:, 0] = weight[:, 0] + alpha * do_error * np.transpose(
phi_s_a_tau)
curr_a = next_a
t += 1
W = weight[:, 0]
return W