def combine_model(path_to_dqn='model_cpu'):
model2 = DQN()
model2.load_state_dict(torch.load(path_to_dqn))
model1 = QModel()
model1.q_function = np.loadtxt('model.csv', delimiter=',')
new_s = game.GameState()
start = time.time()
reward = 0
speed = 0.0
t = 0
image_data, reward, terminate, (x,
y), up, red, _, pedes = new_s.frame_step(0)
last_screen = get_roi(x, y, up)
current_screen = get_roi(x, y, up)
state2 = current_screen - last_screen
while not terminate:
state1 = get_state(image_data, x, y, up, red, pedes)
action = select_action(model1, model2, state1, state2, algorithm=3)
image_data, r, terminate, (
x, y), up, red, sp, pedes = new_s.frame_step(action)
last_screen = current_screen
current_screen = get_roi(x, y, up)
state2 = current_screen - last_screen
reward += r
speed += sp
t += 1
cur = time.time()
print('The game last for {} frames'.format(t))
print('The game last for {} second'.format(cur - start))
print('The total award : {}'.format(reward))
print('The average speed is : {}'.format(speed / t))
cos_ann = True ann_cyc = 5 schedule = Schedule(t0, t1, e0, e1, decay_fun, cosine_annealing=cos_ann, annealing_cycles=ann_cyc) # Policy policy = EpsilonGreedyPolicy(schedule=schedule, value_function=Q) # Reward Function reward_fun = rf_info2d_pos # Action Pre/Post-Processing Action act_fun = act_disc2cont # Agent lr = 1e-4 gamma = 0.99 doubleQ = True # Run doubleQ-DQN sampling from Q_target and bootstraping from Q rb = False rb_max_size = 1e6 rb_batch_size = 64 tau = 0.1 agent = DQN(policy, act_fun, Q, Q_target, state_dim, action_dim, gamma, doubleQ, reward_fun=reward_fun, replay_buffer=rb, max_buffer_size=rb_max_size, batch_size=rb_batch_size, tau=tau, lr=lr) # Training show = False train_agent(agent, desc, file_name, runs, episodes, time_steps, test_episodes, init_state, init_noise, show=show)
import sys
sys.path.append('../src/')
from random_graph import RandGraph
from deep_q_learning import DQN
import matplotlib.pyplot as plt
import pylab
g = RandGraph(graph_type='simple', actors=1000)
h = RandGraph(graph_type='simple', actors=1000)
dqn_g = DQN(g)
dqn_h = DQN(h)
dqn_h.load()
plt.style.use('seaborn-talk')
for i in range(200):
pylab.clf()
plt.subplot(121)
plt.title('Random action')
rand_act = dqn_g.action_space.sample()
a1 = dqn_g.action_space.get_nodes(rand_act)
g.action(a1)
g.plot()
plt.subplot(122)
plt.title('Deep Q Learning')
state = h.get_loading()
a2 = dqn_h.predict(state)
h.action(a2)