def test_online_train_dqn_lunar():
TRAIN_TIMES = 1
TEST_TIMES = 10
BATCH_SIZE = 64
GAMMA = 0.99
EPSILON_THRES = 0.05
TEST_EVERY_EPISODE = 100
REPLAY_MEMORY_SIZE = 100000
NUM_EPISODES = 701
LEARNING_START_EPISODES = 10
TARGET_UPDATE_EVERY_EPISODE = 2
VERBOSE = False
PLOT = False
model_str = 'dqn'
env_str = "lunar"
test_rewards_ave, test_durations_ave = train_main(
model_str,
env_str,
TRAIN_TIMES,
TEST_TIMES,
BATCH_SIZE,
GAMMA,
TEST_EVERY_EPISODE,
REPLAY_MEMORY_SIZE,
NUM_EPISODES,
LEARNING_START_EPISODES,
EPSILON_THRES,
TARGET_UPDATE_EVERY_EPISODE,
VERBOSE,
PLOT,
)
assert np.sum(np.array(test_rewards_ave) > 190) > 0
def test_online_train_lstm_rnn_gridworld():
TRAIN_TIMES = 1
TEST_TIMES = 1
BATCH_SIZE = 4
GAMMA = 0.9
EPSILON_THRES = 0.4
TEST_EVERY_EPISODE = 10
REPLAY_MEMORY_SIZE = 20000
NUM_EPISODES = 5001
LEARNING_START_EPISODES = 500
TARGET_UPDATE_EVERY_EPISODE = 2
VERBOSE = True
PLOT = False
model_str = 'lstm'
env_str = "rnn"
test_rewards_ave, test_durations_ave = train_main(
model_str,
env_str,
TRAIN_TIMES,
TEST_TIMES,
BATCH_SIZE,
GAMMA,
TEST_EVERY_EPISODE,
REPLAY_MEMORY_SIZE,
NUM_EPISODES,
LEARNING_START_EPISODES,
EPSILON_THRES,
TARGET_UPDATE_EVERY_EPISODE,
VERBOSE,
PLOT,
)
# at least 5 in last 10 tests are optimal
assert np.sum(test_rewards_ave[-10:] > 28) > 5
def test_online_train_dqn_cartpole():
TRAIN_TIMES = 1
TEST_TIMES = 20
BATCH_SIZE = 64
GAMMA = 0.9
EPSILON_THRES = 0.05
TEST_EVERY_EPISODE = 20
REPLAY_MEMORY_SIZE = 20000
NUM_EPISODES = 1001
LEARNING_START_EPISODES = 500
TARGET_UPDATE_EVERY_EPISODE = 2
VERBOSE = False
PLOT = False
model_str = 'dqn'
env_str = "cartpole"
test_rewards_ave, test_durations_ave = train_main(
model_str,
env_str,
TRAIN_TIMES,
TEST_TIMES,
BATCH_SIZE,
GAMMA,
TEST_EVERY_EPISODE,
REPLAY_MEMORY_SIZE,
NUM_EPISODES,
LEARNING_START_EPISODES,
EPSILON_THRES,
TARGET_UPDATE_EVERY_EPISODE,
VERBOSE,
PLOT,
)
# cartpole is a little special, we only need at least 1 test optimal
assert np.sum(np.array(test_rewards_ave) > 190) > 0