def __init__(self,
temperature=1.,
with_critic=False,
linear_annealing_kwargs=None):
"""Initializes SoftmaxAgent.
Args:
temperature (float): Softmax temperature parameter.
with_critic (bool): Run the Actor-Critic agent with a value network.
linear_annealing_kwargs (dict): Temperature linear annealing
schedule with keys: 'max_value', 'min_value', 'n_epochs',
overrides temperature. If None, then uses constant temperature
value.
"""
if linear_annealing_kwargs is not None:
parameter_schedules = {
'distribution.temperature':
schedules.LinearAnnealing(**linear_annealing_kwargs)
}
else:
parameter_schedules = {}
super().__init__(distribution=distributions.CategoricalDistribution(
temperature=temperature),
with_critic=with_critic,
parameter_schedules=parameter_schedules)
def __init__(self,
epsilon=.05,
with_critic=False,
linear_annealing_kwargs=None):
"""Initializes EpsilonGreedyAgent.
Args:
epsilon (float): Probability of taking random action.
with_critic (bool): Run the Actor-Critic agent with a value network.
linear_annealing_kwargs (dict): Epsilon linear annealing
schedule with keys: 'max_value', 'min_value', 'n_epochs',
overrides epsilon. If None, then uses constant epsilon value.
"""
if linear_annealing_kwargs is not None:
parameter_schedules = {
'distribution.epsilon':
schedules.LinearAnnealing(**linear_annealing_kwargs)
}
else:
parameter_schedules = {}
super().__init__(distribution=distributions.EpsilonGreedyDistribution(
epsilon=epsilon),
with_critic=with_critic,
parameter_schedules=parameter_schedules)
def test_linear_annealing_schedule(mock_env):
# Set up
attr_name = 'pied_piper'
param_values = list(range(10, 0, -1))
max_value = max(param_values)
min_value = min(param_values)
n_epochs = len(param_values)
agent = core.RandomAgent(parameter_schedules={
attr_name: schedules.LinearAnnealing(max_value, min_value, n_epochs)
})
# Run & Test
for epoch, x_value in enumerate(param_values):
testing.run_without_suspensions(agent.solve(mock_env, epoch=epoch))
assert getattr(agent, attr_name) == x_value