def __init__(self, history_mgr=None, *args, **kwargs):
self.args = args
self.kwargs = kwargs
# check if the object is responsible to keep the history of its interaction
if isinstance(history_mgr, grl.HistoryManager):
# an external history manager is provided, so no need to keep history in the object
self.keep_history = False
self.hm = history_mgr
else:
# the object will keep its own history
self.keep_history = True
# the initiated history manager uses the internal state_func to get states from histories
self.hm = grl.HistoryManager(maxlen=self.kwargs.get(
'max_history', None),
state_map=self.state_func)
self.sm = grl.StateManager(self.transition_func)
self.am = grl.ActionManager()
self.pm = grl.PerceptManager(self.emission_func)
self.rm = grl.RewardManager(self.reward_func)
# the order is important if the objects are interacting in a sequence
self.order = self.kwargs.get('order', math.nan)
# call the setup function for any object-dependent user-specified configurations
# from a user perspective, setup() is the entry point of an object
# the user should not use the internal __init__
self.setup()
def __init__(self, history_mgr=None, *args, **kwargs):
self.args = args
self.kwargs = kwargs
# If the history_mgr belongs to the HistoryManager class then
# store the history_mgr as hm, else create a new HistoryManager instance
if isinstance(history_mgr, grl.HistoryManager):
self.keep_history = False
self.hm = history_mgr
else:
self.keep_history = True
self.hm = grl.HistoryManager(maxlen=self.kwargs.get(
'max_history', None),
state_map=self.state_func)
self.sm = grl.StateManager(self.transition_func)
self.am = grl.ActionManager()
self.pm = grl.PerceptManager(self.emission_func)
self.rm = grl.RewardManager(self.reward_func)
self.order = self.kwargs.get('order', math.nan)
self.setup()
def phi_extreme_q(h, *args, **kwargs):
eps = kwargs.get('eps', 0.1)
q_func = kwargs.get('q_func', None)
q = q_func(h, *args, **kwargs)
q = q // eps
s = tuple(q[k] for k in sorted(q))
return s
def phi_last_percept(h, *args, **kwargs):
# extract last percept
return h[-1]
history_mgr = grl.HistoryManager(maxlen=10, state_map=phi_extreme_va)
#domain = BlindMaze(history_mgr, maze_len=2)
#domain = SimpleMDP(history_mgr)
domain = SlipperyHill(history_mgr)
#domain = DynamicKeys(history_mgr)
#agent = RandomAgent(history_mgr)
agent = FrequencyAgent(history_mgr,
exploration_factor=0.1,
discount_factor=0.999)
#agent = GreedyQAgent(history_mgr, value_function_persist=False, exploration_factor=0.3)
# o_domain = domain
# domain = grl.BinaryMock(history_mgr)
# domain.hook(o_domain)
agent.interact(domain)
def setup(self):
self.am.action_space = [0, 1]
self.r_dummy = self.kwargs.get('r_dummy', 0.0)
self.hm_ae = grl.HistoryManager(maxlen=self.hm.maxlen)
self.domain = None
self.restrict_A_cache = dict()
def __init__(self, reward_func=lambda h, *x, **y: 0):
self.reward_func = reward_func
self.hm = grl.HistoryManager()
