def _sample_batch(self, batch_size):
obs_batch, act_batch, reward_batch, next_obs_batch, is_done_batch, next_act_batch = \
self.exp_buffer.sample(batch_size)
obs_batch = min_max_scaling(obs_batch)
next_obs_batch = min_max_scaling(next_obs_batch)
is_done_batch = is_done_batch.astype("float32")
reward_batch = reward_batch.astype("float32")
return {"obs": obs_batch, "actions": act_batch, "rewards": reward_batch,
"next_obs": next_obs_batch, "is_done": is_done_batch, "next_actions": next_act_batch}
def _sample_batch(self, batch_size):
obs_batch, act_batch, reward_batch, next_obs_batch, is_done_batch = self.exp_buffer.sample(
batch_size)
obs_batch = obs_batch.reshape(obs_batch.shape[0], 1,
obs_batch.shape[1] * obs_batch.shape[2])
next_obs_batch = next_obs_batch.reshape(
next_obs_batch.shape[0], 1,
next_obs_batch.shape[1] * next_obs_batch.shape[2])
obs_batch = min_max_scaling(obs_batch)
next_obs_batch = min_max_scaling(next_obs_batch)
is_done_batch = is_done_batch.astype("float32")
reward_batch = reward_batch.astype("float32")
return {
"obs": obs_batch,
"actions": act_batch,
"rewards": reward_batch,
"next_obs": next_obs_batch,
"is_done": is_done_batch
}
def decision(self, state_space: np.ndarray, action_space: ActionSpace):
"""
triggered by get play turn method of super class.
This is the method were the magic should happen that chooses the right action
:param state_space:
:param action_space:
:return:
"""
# preprocess state space
# normalizing state space between zero and one
state_space = min_max_scaling(state_space)
state_space = state_space.reshape(1, multiply(*state_space.shape))
qvalues = self._get_qvalues([state_space])
decision = self._sample_actions(qvalues, action_space)
return decision