def __init__(self, environment, configuration, name='', debug=False):
num_actions = environment.action_space.n
num_states = get_num_states(environment)
self.Q = np.zeros((num_states, num_actions))
# init super lastly, because it starts the training
super().__init__(environment=environment, debug=debug, name=name, configuration=configuration)
def __init__(self,
environment,
memory_size,
nb_variational_circuits,
configuration,
name='',
debug=False):
self.D = [] # Memory
self.memory_size = memory_size
self.nb_variational_circuits = nb_variational_circuits
self.debug = debug
num_actions = environment.action_space.n
num_states = get_num_states(environment)
self.nb_qbits = int(np.max([num_actions, np.log2(num_states)]))
self.theta = np.array([
2 * np.pi * np.random.random(3)
for j in range(self.nb_qbits * self.nb_variational_circuits)
]).flatten()
self.init_memory(environment)
# init super lastly, because it starts the training
super().__init__(environment=environment,
debug=debug,
name=name,
configuration=configuration)
def __init__(self, environment, configuration, name, debug=False):
self.environment = environment
self.debug = debug
self.name = name
self.configuration = configuration
self.num_actions = environment.action_space.n
self.num_states = get_num_states(environment)
def __init__(self, environment, configuration, name='', debug=False):
self.D = [] # Memory
self.num_states = get_num_states(environment)
self.model = configuration.model
self.model.compile(loss='mean_squared_error', optimizer='sgd')
self.target_model = clone_model(self.model)
self.memory_size = configuration.memory_size
self.train_counter = 0
self.init_memory(environment)
# init super lastly, because it starts the training
super().__init__(environment=environment, debug=debug, name=name, configuration=configuration)
def __init__(self, environment, configuration, name='', debug=False):
# init super
Agent.__init__(self,
environment=environment,
debug=debug,
name=name,
configuration=configuration)
self.D = [] # Memory
self.num_states = get_num_states(environment)
self.model = configuration.model
self.target_model = configuration.clone_model(self.model)
self.memory_size = configuration.memory_size
self.train_counter = 0
self.episode = 0
self.init_memory(environment)
def __init__(self, environment, configuration, name='', verbosity_level = 10, debug=False):
self.D = [] # Memory
self.memory_size = configuration.memory_size
self.nb_variational_circuits = configuration.nb_variational_circuits
self.debug = debug
self.train_counter = 0
self.episode = 0
num_actions = environment.action_space.n
num_states = get_num_states(environment)
self.nb_circuit_params = 1
self.nb_qbits = int(np.max([num_actions, np.log2(num_states)]))
self.theta = np.array(
[np.random.random(self.nb_circuit_params) for j in range(self.nb_qbits * self.nb_variational_circuits)]).flatten()
self.target_theta = copy.deepcopy(self.theta.copy())
self.init_memory(environment)
self.loss_mem = []
# init super lastly, because it starts the training
super().__init__(environment=environment, debug=debug, name=name, configuration=configuration)
# target update
target_replacement = 10
#alpha, gamma, epsilon
training_params_Q = [.6, .8, .9]
# Instantiate the trainer
confQ = Configuration(nb_iterations=nb_iterations, training_params=training_params_Q, cooling_scheme=cooling_schemes,
batch_size=1, average=int(batch_size/100), test_steps=1, verbosity_level=1e20)
#alpha, gamma, epsilon
training_params_DQN = [.6, .8, .9] # DQN opt .1, .8, .9 overfitting nach 590
# NN model for DQN
num_states = get_num_states(env)
model = Sequential()
model.add(Dense(8, input_shape=(16, ), activation='tanh'))
model.add(Dense(env.action_space.n, activation='linear'))
model.compile(loss=Huber(), optimizer=SGD(training_params_DQN[0]))
confDQN = Configuration(nb_iterations=nb_iterations, training_params=training_params_DQN[1:],
cooling_scheme=cooling_schemes[1:], batch_size=batch_size,
memory_size=memory_size, average=int(nb_iterations/10), model=model,
target_replacement=target_replacement, test_steps=1, verbosity_level=10)
confDQN.clone_model = clone_model
confDQN.embedding = lambda state: to_categorical(state, num_classes=num_states).reshape(num_states)
training_params_VQDQN = [.22, .8, .9] # DQN opt .1, .8, .9 overfitting nach 590
confVQD = Configuration(nb_iterations=nb_iterations, training_params=training_params_VQDQN, cooling_scheme=cooling_schemes,