def dist_fn(dist):
try:
greedy_action = dist.mode()
except NotImplementedError:
raise ValueError("Your network's distribution does not implement "
"mode making it incompatible with a greedy policy.")
return greedy_policy.DeterministicWithLogProb(loc=greedy_action)
def deterministic_action_distribution(self, time_step):
"""
Produce a deterministic tfp.distribution centered on the action
from the current table.
"""
obs = time_step.observation
action = self.table_lookup(obs)
return greedy_policy.DeterministicWithLogProb(loc=action)