def __init__(self,
player_id,
num_actions,
state_representation_size,
step_size=0.5,
epsilon_schedule=rl_tools.ConstantSchedule(0.2),
discount_factor=1.0,
initialization='random',
name='linear_qlearner'):
"""Initialize the Q-Learning agent."""
self._player_id = player_id
self._num_actions = num_actions
self._state_representation_size = state_representation_size
self._step_size = step_size
self._epsilon_schedule = epsilon_schedule
self._epsilon = epsilon_schedule.value
self._discount_factor = discount_factor
# random init state features + one hot encoded actions for (S,a)
if initialization == 'random':
self._weights = np.random.random(self._state_representation_size * self._num_actions)
elif initialization == 'zero':
self._weights = np.zeros(self._state_representation_size * self._num_actions)
elif type(initialization) is list and len(initialization) == self._state_representation_size * self._num_actions:
self._weights = initialization
else:
raise ValueError("Not implemented, choose initialization from 'random', 'zero' and a custom list.")
self._prev_info_state = None
def create_epsilon_schedule(sched_str):
"""Creates an epsilon schedule from the string as desribed in the flags."""
values = FLAGS.epsilon_schedule.split(",")
if values[0] == "linear":
assert len(values) == 4
return rl_tools.LinearSchedule(float(values[1]), float(values[2]),
int(values[3]))
elif values[0] == "constant":
assert len(values) == 2
return rl_tools.ConstantSchedule(float(values[1]))
else:
print("Unrecognized schedule string: {}".format(sched_str))
sys.exit()
def __init__(self,
player_id,
num_actions,
step_size=0.1,
epsilon_schedule=rl_tools.ConstantSchedule(0.2),
discount_factor=1.0):
"""Initialize the Q-Learning agent."""
self._player_id = player_id
self._num_actions = num_actions
self._step_size = step_size
self._epsilon_schedule = epsilon_schedule
self._epsilon = epsilon_schedule.value
self._discount_factor = discount_factor
self._q_values = collections.defaultdict(
lambda: collections.defaultdict(float))
self._prev_info_state = None
def __init__(self,
player_id,
num_actions,
step_size=0.1,
epsilon_schedule=rl_tools.ConstantSchedule(0.2),
discount_factor=1.0,
centralized=False):
"""Initialize the Q-Learning agent."""
self._player_id = player_id
self._num_actions = num_actions
self._step_size = step_size
self._epsilon_schedule = epsilon_schedule
self._epsilon = epsilon_schedule.value
self._discount_factor = discount_factor
self._centralized = centralized
self._q_values = collections.defaultdict(valuedict)
self._prev_info_state = None
self._last_loss_value = None