def learning(episode_count):
env = Env()
actions = 8
input_dim = (2, 84, 84)
scores = []
avg_scores = deque([])
gamma = 0.9
lr = 0.0001
exp_buffer_size = 25000
batch_size = 256
buffer = ExperienceReplay(exp_buffer_size, input_dim)
agent = AgentDDQN(gamma,
actions,
Net(),
buffer,
lr,
update_steps=1000,
batch_size=batch_size,
path="trained_models/sc2_beacon",
epsilon_dec=1e-5)
d_now = datetime.datetime.now()
for i in range(0, episode_count):
score = 0
terminal = False
state = env.reset()
state = torch.from_numpy(state).double()
while not terminal:
action = agent.choose_action(state)
state_, reward, terminal = env.step(action)
state_ = torch.from_numpy(state_).double()
score += reward
agent.store(state, action, reward, state_, terminal)
agent.learn()
state = state_
avg_scores.append(score)
if len(avg_scores) > MAX_SCORE_COUNT:
avg_scores.popleft()
scores.append(score)
print('episode: ', i, '\t\tscore: ',
+round(score, 3), '\t\tcumulative score:',
round(np.average(scores), 3), '\t\tavg-100 score:',
round(np.average(avg_scores), 3), '\t\tepsilon: ',
round(agent.epsilon, 4))
if i % 100 == 0:
d_end = datetime.datetime.now()
d = d_end - d_now
print('time: ', d)
if i % 100 == 0:
agent.save_model()
def animation():
env = Env()
actions = 2 * 64 * 64
input_dim = (3, 64, 64)
scores = []
avg_scores = deque([])
gamma = 0.9
lr = 0.0001
exp_buffer_size = 0
batch_size = 0
buffer = ExperienceReplay(exp_buffer_size, input_dim)
agent = AgentDDQN(gamma,
actions,
Net(),
buffer,
lr,
update_steps=1000,
batch_size=batch_size,
epsilon=0.01,
path="trained_models/sc2_collect_shards")
agent.load_model()
i = 0
while True:
score = 0
terminal = False
state = env.reset()
state = torch.from_numpy(state).double()
i += 1
while not terminal:
action = agent.choose_action(state)
state_, reward, terminal = env.step(action)
state = torch.from_numpy(state_).double()
score += reward
avg_scores.append(score)
if len(avg_scores) > MAX_SCORE_COUNT:
avg_scores.popleft()
scores.append(score)
print('episode: ', i, '\t\tcumulative score:',
round(np.average(scores), 3), '\t\tavg-100 score:',
round(np.average(avg_scores), 3))
def learning():
env = gym.make('LunarLander-v2')
actions = 4
state_dim = (8, )
games_count = 1000
scores = []
experience_replay = ExperienceReplay(10000, state_dim)
agent = AgentDDQN(0.99, actions, Net(), experience_replay, 0.001)
d_now = datetime.datetime.now()
for i in range(1, games_count):
score = 0
terminal = False
state = env.reset()
state = torch.from_numpy(state).double()
while not terminal:
action = agent.choose_action(state)
state_, reward, terminal, _ = env.step(action)
state_ = torch.from_numpy(state_).double()
agent.store(state, action, reward, state_, terminal)
agent.learn()
state = state_
score += reward
scores.append(score)
print('episode: ', i, '\t\tscore: ', +score, '\t\taverage score:',
np.average(scores[-100:]), '\t\tepsilon: ', agent.epsilon)
if i % 10 == 0:
d_end = datetime.datetime.now()
d = d_end - d_now
print('time: ', d)
def learning(num):
env = make_env('PongNoFrameskip-v4')
actions = 6
input_dim = (4, 80, 80)
scores = []
experience_replay = ExperienceReplay(25000, input_dim)
agent = AgentDDQN(0.99,
actions,
Net(),
experience_replay,
0.0001,
update_steps=1000,
batch_size=32,
epsilon_min=0.02,
epsilon=1.0,
epsilon_dec=1e-4)
'''
d_now = datetime.datetime.now()
text = 'episode,score,time,step,epsilon'
for i in range(1, games_count):
score = 0
terminal = False
state = env.reset()
state = torch.from_numpy(state).double()
while not terminal:
action = agent.choose_action(state)
state_, reward, terminal, _ = env.step(action)
state_ = torch.from_numpy(state_).double()
agent.store(state, action, reward, state_, terminal)
agent.learn()
state = state_
score += reward
scores.append(score)
print('episode: ', i, '\t\tscore: ', + score, '\t\taverage score:' , np.average(scores[-100:]), '\t\tepsilon: ', agent.epsilon)
if i % 10 == 0:
d_end = datetime.datetime.now()
d = d_end - d_now
print('time: ', d)
'''
i = 0
step = 0
d_start = datetime.datetime.now()
print('start ', d_start)
d_end = datetime.datetime.now() + datetime.timedelta(minutes=180)
text = 'episode,score,time,step,epsilon'
while d_end > datetime.datetime.now():
score = 0
terminal = False
state = env.reset()
state = torch.from_numpy(state).double()
i += 1
while not terminal:
step += 1
action = agent.choose_action(state)
state_, reward, terminal, _ = env.step(action)
state_ = torch.from_numpy(state_).double()
score += reward
agent.store(state, action, reward, state_, terminal)
agent.learn()
state = state_
scores.append(score)
avg = np.average(scores[-100:])
d = datetime.datetime.now() - d_start
text += '\n' + str(i) + ',' + str(avg) + ',' + str(d) + ',' + str(
step) + ',' + str(agent.epsilon)
write_to_file(text, 'logs/pong/log_1_' + str(num) + '_DDQN.txt')