def from_discrete_action_model_parameters(
cls, params: rlp.DiscreteActionModelParameters):
return cls(
actions=params.actions,
rl=params.rl,
double_q_learning=params.rainbow.double_q_learning,
num_atoms=params.rainbow.num_atoms,
minibatch_size=params.training.minibatch_size,
minibatches_per_step=params.training.minibatches_per_step,
cpe_optimizer=rlp.OptimizerParameters(
optimizer=params.training.optimizer,
learning_rate=params.training.learning_rate,
l2_decay=params.training.l2_decay,
),
optimizer=rlp.OptimizerParameters(
optimizer=params.training.optimizer,
learning_rate=params.training.learning_rate,
l2_decay=params.rainbow.c51_l2_decay,
),
evaluation=params.evaluation,
)
def from_discrete_action_model_parameters(
cls, params: DiscreteActionModelParameters):
return cls(
actions=params.actions,
rl=params.rl,
double_q_learning=params.rainbow.double_q_learning,
bcq=BCQConfig(drop_threshold=params.rainbow.bcq_drop_threshold)
if params.rainbow.bcq else None,
minibatch_size=params.training.minibatch_size,
minibatches_per_step=params.training.minibatches_per_step,
optimizer=rlp.OptimizerParameters(
optimizer=params.training.optimizer,
learning_rate=params.training.learning_rate,
l2_decay=params.training.l2_decay,
),
evaluation=params.evaluation,
)