def __init__(self, config):
config.default(PolicyGradientModel.default_config)
# distribution
self.distribution = dict()
for name, action in config.actions:
if 'distribution' in action:
kwargs = dict(action)
self.distribution[name] = Distribution.from_config(
config=action.distribution, kwargs=kwargs)
elif action.continuous:
if action.min_value is None:
assert action.max_value is None
self.distribution[name] = Gaussian(shape=action.shape)
else:
assert action.max_value is not None
self.distribution[name] = Beta(min_value=action.min_value,
max_value=action.max_value,
shape=action.shape)
else:
self.distribution[name] = Categorical(
shape=action.shape, num_actions=action.num_actions)
# baseline
if config.baseline is None:
self.baseline = None
else:
self.baseline = Baseline.from_config(config=config.baseline)
# advantage estimation
self.gae_rewards = config.gae_rewards
self.gae_lambda = config.gae_lambda
self.normalize_rewards = config.normalize_rewards
super(PolicyGradientModel, self).__init__(config)
def __init__(self, config):
config.default(PolicyGradientModel.default_config)
# distribution
self.distribution = dict()
for name, action in config.actions:
if 'distribution' in action:
if not action.continuous:
kwargs = dict(num_actions=action.num_actions)
elif 'min_value' in action:
kwargs = dict(min_value=action.min_value, max_value=action.max_value)
else:
kwargs = dict()
self.distribution[name] = Distribution.from_config(config=action.distribution, kwargs=kwargs)
# elif 'min_value' in action:
# ...
elif action.continuous:
self.distribution[name] = Gaussian()
else:
self.distribution[name] = Categorical(num_actions=action.num_actions)
# baseline
if config.baseline is None:
self.baseline = None
else:
self.baseline = Baseline.from_config(config=config.baseline)
super(PolicyGradientModel, self).__init__(config)
# advantage estimation
self.generalized_advantage_estimation = config.generalized_advantage_estimation
if self.generalized_advantage_estimation:
self.gae_lambda = config.gae_lambda
self.normalize_advantage = config.normalize_advantage
