def __init__(self, config):
"""Initializes the reinforcement learning agent.
Args:
config (Configuration): configuration object containing at least `states`, `actions`, `preprocessing` and
'exploration`.
"""
assert self.__class__.name is not None and self.__class__.model is not None
config.default(Agent.default_config)
# states config and preprocessing
self.preprocessing = dict()
if 'type' in config.states:
# only one state
config.states = dict(state=config.states)
self.unique_state = True
if config.preprocessing is not None:
config.preprocessing = dict(state=config.preprocessing)
else:
self.unique_state = False
for name, state in config.states:
if config.preprocessing is not None and name in config.preprocessing:
preprocessing = Preprocessing.from_config(
config=config.preprocessing[name])
self.preprocessing[name] = preprocessing
state.shape = preprocessing.processed_shape(shape=state.shape)
# actions config and exploration
self.continuous_actions = list()
self.exploration = dict()
if 'continuous' in config.actions:
# only one action
if config.actions.continuous:
self.continuous_actions.append('action')
config.actions = dict(action=config.actions)
if config.exploration is not None:
config.exploration = dict(action=config.exploration)
self.unique_action = True
else:
self.unique_action = False
for name, action in config.actions:
if action.continuous:
self.continuous_actions.append(name)
if config.exploration is not None and name in config.exploration:
self.exploration[name] = Exploration.from_config(
config=config.exploration[name])
self.states_config = config.states
self.actions_config = config.actions
self.model = self.__class__.model(config)
self.episode = 0
self.timestep = 0
# Reset internal state - needs to be called after every episode
self.next_internal = self.current_internal = self.model.reset()
for preprocessing in self.preprocessing.values():
preprocessing.reset()
def __init__(self, config):
"""Initializes the reinforcement learning agent.
Args:
config (Configuration): configuration object containing at least `states`, `actions`, `preprocessing` and
'exploration`.
"""
assert self.__class__.name is not None and self.__class__.model is not None
config.default(Agent.default_config)
# states config and preprocessing
self.preprocessing = dict()
if 'type' in config.states:
# only one state
config.states = dict(state=config.states)
self.unique_state = True
if config.preprocessing is not None:
config.preprocessing = dict(state=config.preprocessing)
else:
self.unique_state = False
for name, state in config.states:
state.default(dict(type='float'))
if isinstance(state.shape, int):
state.shape = (state.shape, )
if config.preprocessing is not None and name in config.preprocessing:
preprocessing = Preprocessing.from_config(
config=config.preprocessing[name])
self.preprocessing[name] = preprocessing
state.shape = preprocessing.processed_shape(shape=state.shape)
# actions config and exploration
self.exploration = dict()
if 'continuous' in config.actions:
# only one action
config.actions = dict(action=config.actions)
if config.exploration is not None:
config.exploration = dict(action=config.exploration)
self.unique_action = True
else:
self.unique_action = False
for name, action in config.actions:
if action.continuous:
action.default(dict(shape=(), min_value=None, max_value=None))
else:
action.default(dict(shape=()))
if isinstance(action.shape, int):
action.shape = (action.shape, )
if config.exploration is not None and name in config.exploration:
self.exploration[name] = Exploration.from_config(
config=config.exploration[name])
self.states_config = config.states
self.actions_config = config.actions
self.model = self.__class__.model(config)
self.episode = -1
self.timestep = 0
self.reset()
def __init__(self, config):
"""Initializes the reinforcement learning agent.
Args:
config (Configuration): configuration object containing at least `states`, `actions`, `preprocessing` and
'exploration`.
"""
assert self.__class__.name is not None and self.__class__.model is not None
config.default(Agent.default_config)
# states config and preprocessing
self.preprocessing = dict()
if 'type' in config.states:
# only one state
config.states = dict(state=config.states)
self.unique_state = True
if config.preprocessing is not None:
config.preprocessing = dict(state=config.preprocessing)
else:
self.unique_state = False
for name, state in config.states:
if config.preprocessing is not None and name in config.preprocessing:
preprocessing = Preprocessing.from_config(config=config.preprocessing[name])
self.preprocessing[name] = preprocessing
state.shape = preprocessing.processed_shape(shape=state.shape)
# actions config and exploration
self.continuous_actions = list()
self.exploration = dict()
if 'continuous' in config.actions:
# only one action
if config.actions.continuous:
self.continuous_actions.append('action')
config.actions = dict(action=config.actions)
if config.exploration is not None:
config.exploration = dict(action=config.exploration)
self.unique_action = True
else:
self.unique_action = False
for name, action in config.actions:
if action.continuous:
self.continuous_actions.append(name)
if config.exploration is not None and name in config.exploration:
self.exploration[name] = Exploration.from_config(config=config.exploration[name])
self.states_config = config.states
self.actions_config = config.actions
self.model = self.__class__.model(config)
self.episode = 0
self.timestep = 0
# Reset internal state - needs to be called after every episode
self.next_internal = self.current_internal = self.model.reset()
for preprocessing in self.preprocessing.values():
preprocessing.reset()
def __init__(self, config, model=None):
"""Initializes the reinforcement learning agent.
Args:
config (Configuration): configuration object containing at least `states`, `actions`, `preprocessing` and
'exploration`.
model (Model): optional model instance. If not supplied, a new model is created.
"""
assert self.__class__.name is not None and self.__class__.model is not None
config.default(Agent.default_config)
self.logger = logging.getLogger(__name__)
self.logger.setLevel(util.log_levels[config.log_level])
# states config and preprocessing
self.preprocessing = dict()
if 'type' in config.states:
# only one state
config.states = dict(state=config.states)
self.unique_state = True
if config.preprocessing is not None:
config.preprocessing = dict(state=config.preprocessing)
else:
self.unique_state = False
for name, state in config.states:
state.default(dict(type='float'))
if isinstance(state.shape, int):
state.shape = (state.shape, )
if config.preprocessing is not None and name in config.preprocessing:
preprocessing = Preprocessing.from_config(
config=config.preprocessing[name])
self.preprocessing[name] = preprocessing
state.shape = preprocessing.processed_shape(shape=state.shape)
# actions config and exploration
self.exploration = dict()
if 'continuous' in config.actions:
# only one action
config.actions = dict(action=config.actions)
if config.exploration is not None:
config.exploration = dict(action=config.exploration)
self.unique_action = True
else:
self.unique_action = False
for name, action in config.actions:
if action.continuous:
action.default(dict(shape=(), min_value=None, max_value=None))
else:
action.default(dict(shape=()))
if isinstance(action.shape, int):
action.shape = (action.shape, )
if config.exploration is not None and name in config.exploration:
self.exploration[name] = Exploration.from_config(
config=config.exploration[name])
self.states_config = config.states
self.actions_config = config.actions
if model is None:
self.model = self.__class__.model(config)
else:
if not isinstance(model, self.__class__.model):
raise TensorForceError(
"Supplied model class `{}` does not match expected agent model class `{}`"
.format(
type(model).__name__, self.__class__.model.__name__))
self.model = model
not_accessed = config.not_accessed()
if not_accessed:
self.logger.warning("Configuration values not accessed: {}".format(
', '.join(not_accessed)))
self.episode = -1
self.timestep = 0
self.reset()