# Initialize the loss function
loss_function = tf.keras.losses.MeanSquaredError()
# Initialize the optimizer
optimizer = tf.keras.optimizers.Adam(learning_rate)
# Initialize
ray.init(log_to_driver=False)
manager = SampleManager(**kwargs)
# Where to save your results to: create this directory in advance!
saving_path = os.getcwd() + "/progress_LunarLander"
# Initialize buffer
manager.initilize_buffer(buffer_size)
# Fill buffer
manager.store_in_buffer(manager.get_data(total_steps=buffer_size))
# Initialize progress aggregator
manager.initialize_aggregator(
path=saving_path, saving_after=5, aggregator_keys=["loss", 'reward', 'time']
)
rewards = []
# Get initial agent
agent = manager.get_agent()
print('TRAINING')
"action_sampling_type": "epsilon_greedy",
"num_episodes": 20,
"epsilon": epsilon,
}
ray.init(log_to_driver=False)
manager = SampleManager(**kwargs)
# where to save your results to: create this directory in advance!
saving_path = os.getcwd() + "/progress_cartpole"
# keys for replay buffer -> what you will need for optimization
optim_keys = ["state", "action", "reward", "state_new", "not_done"]
# initialize buffer
manager.initilize_buffer(buffer_size, optim_keys)
# initilize progress aggregator
manager.initialize_aggregator(path=saving_path,
saving_after=5,
aggregator_keys=["loss", "time_steps"])
# initial testing:
print("test before training: ")
manager.test(test_steps, test_episodes=10, do_print=True, render=True)
# get initial agent
agent = manager.get_agent()
for e in range(epochs):
'model_kwargs':model_kwargs,
"action_sampling_type": "epsilon_greedy",
"epsilon": EPSILON
}
manager = SampleManager(**kwargs)
# specify where to save results and ensure that the folder exists
saving_path = Path(os.getcwd() + SAVING_DIRECTORY)
saving_path.mkdir(parents=True, exist_ok=True)
saving_path_model = Path(os.getcwd() + SAVING_DIRECTORY + '/model')
saving_path_model.mkdir(parents=True, exist_ok=True)
# initialize manager
optim_keys = ['state', 'action', 'reward', 'state_new', 'not_done']
manager.initilize_buffer(BUFFER_SIZE, optim_keys)
aggregator_keys=['loss', 'time_steps', 'reward']
manager.initialize_aggregator(saving_path, 5, aggregator_keys)
# initialize the optimizers
optimizer = Adam(learning_rate=LEARNING_RATE)
print('# =============== INITIAL TESTING =============== #')
manager.test(MAX_TEST_STEPS, 5, evaluation_measure='time_and_reward', do_print=True, render=True)
# get the initial agent
agent = manager.get_agent()
print('# =============== START TRAINING ================ #')
for e in range(1, EPOCHS+1):
print(f'# ============== EPOCH {e}/{EPOCHS} ============== #')