def action(self, context):
"""Samples rewards from posterior, and chooses best action accordingly."""
# Round robin until each action has been selected "initial_pulls" times
if self.t < self.hparams.num_actions * self.hparams.initial_pulls:
return self.t % self.hparams.num_actions
vals = []
states, rewards, actions = bandit_utils.get_data_with_masked_rewards(
self.data_h)
state_action_pairs = tf.concat([states, actions], axis=-1)
historical_x = tf.to_float(tf.expand_dims(state_action_pairs, axis=0))
historical_y = tf.to_float(rewards.reshape(1, -1, 1))
context = tf.to_float(context)
tiled_context = tf.concat([
tf.tile(tf.reshape(context, [1, -1]),
[self.hparams.num_actions, 1]), self._one_hot_vectors
],
axis=-1)
target_x = tf.expand_dims(tiled_context, axis=0)
target_y = None
prediction = self.snp(historical_x, historical_y, target_x, target_y)
vals = tf.squeeze(prediction.distribution.mean())
return tf.argmax(vals).numpy()
def action(self, context):
"""Samples rewards from posterior, and chooses best action accordingly.
Args:
context: A d-dimensional np.ndarray with the context.
Returns:
Greedy action based on Thompson sampling.
"""
# Round robin until each action has been selected "initial_pulls" times
if self.t < self.hparams.num_actions * self.hparams.initial_pulls:
return self.t % self.hparams.num_actions
vals = []
context = tf.to_float(context)
if self._is_anp:
contexts, rewards, actions = self.data_h.get_data_with_weights()
historical_x = tf.to_float(tf.expand_dims(contexts, axis=0))
historical_y = tf.to_float(
tf.expand_dims(rewards * actions, axis=0))
target_x = tf.expand_dims(tf.reshape(context, [1, -1]), axis=0)
else:
contexts, rewards, actions = utils.get_data_with_masked_rewards(
self.data_h)
context_action_pairs = tf.concat([contexts, actions], axis=-1)
historical_x = tf.to_float(
tf.expand_dims(context_action_pairs, axis=0))
historical_y = tf.to_float(rewards.reshape(1, -1, 1))
tiled_context = tf.concat([
tf.tile(tf.reshape(context, [1, -1]),
[self.hparams.num_actions, 1]), self._one_hot_vectors
],
axis=-1)
target_x = tf.expand_dims(tiled_context, axis=0)
target_y = None
predictions = self.snp(historical_x, historical_y, target_x, target_y)
vals = tf.squeeze(predictions.distribution.mean())
return tf.argmax(vals).numpy()