class ApproxActionValueFunction(DeepSarsaApproxActionValueFunction):
def __init__(self, handicappers=None):
super(DeepSarsaApproxActionValueFunction, self).__init__()
self._handicappers = handicappers
def setup(self):
self.delegate = MLPOneLayerActionRecordScaledScalarFeaturesValueFunction(NB_UNIT, blind_structure, self._handicappers)
self.delegate.setup()
def construct_features(self, state, action):
return self.delegate.construct_features(state, action)
def approx_predict_value(self, features):
return self.delegate.approx_predict_value(features)
def backup_on_minibatch(self, backup_minibatch):
X = np.array([self.delegate.construct_features(state, action)[0]
for state, action, target in backup_minibatch])
Y_info = [(action, target) for _state, action, target in backup_minibatch]
Y = self.delegate.model.predict_on_batch(X)
assert len(Y) == len(Y_info)
for y, (action, target) in zip(Y, Y_info): y[action_index(action)] = target
loss = self.delegate.model.train_on_batch(X, Y)
self.delegate.loss_history.append(loss)
self.delegate.prediction_cache = (None, None)
def save(self, save_dir_path):
self.delegate.save(save_dir_path)
def load(self, load_dir_path):
self.delegate.load(load_dir_path)
class ApproxActionValueFunction(SarsaApproxActionValueFunction):
def __init__(self, handicappers=None):
super(SarsaApproxActionValueFunction, self).__init__()
self._handicappers = handicappers
def setup(self):
self.delegate = MLPOneLayerActionRecordScaledScalarFeaturesValueFunction(
NB_UNIT, blind_structure, self._handicappers)
self.delegate.setup()
def construct_features(self, state, action):
return self.delegate.construct_features(state, action)
def approx_predict_value(self, features):
return self.delegate.approx_predict_value(features)
def approx_backup(self, features, backup_target, alpha):
self.delegate.approx_backup(features, backup_target, alpha)
def save(self, save_dir_path):
self.delegate.save(save_dir_path)
def load(self, load_dir_path):
self.delegate.load(load_dir_path)