def act(self, step_wrappers):
# """""""""""""""""""""
# Act
# """""""""""""""""""""
SeatActorBase.act_mixed(step_wrappers=step_wrappers,
owner=self.seat_id,
br_learner=self.br_learner,
avg_learner=self.avg_learner,
current_policy_tags=self._current_policy_tags,
random_prob=self.br_learner.eps)
# """""""""""""""""""""
# Add to memories
# """""""""""""""""""""
for sw in step_wrappers:
e_i = sw.env_idx
if (self._current_policy_tags[e_i] == SeatActorBase.BR) and (
self._t_prof.add_random_actions_to_buffer or
(not sw.action_was_random)):
self._avg_buf_savers[e_i].add_step(
pub_obs=sw.obs,
a=sw.action,
legal_actions_mask=rl_util.get_legal_action_mask_np(
n_actions=self._env_bldr.N_ACTIONS,
legal_actions_list=sw.legal_actions_list))
self._br_memory_savers[e_i].add_experience(
obs_t_before_acted=sw.obs,
a_selected_t=sw.action,
legal_actions_list_t=sw.legal_actions_list)
def act_for_avg_opp(self, step_wrappers):
"""
Purely random because that's how it should be for correct reach
"""
SeatActorBase.act_avg(
step_wrappers=step_wrappers,
owner=self.owner,
avg_learner=self.avg_learner,
)
def act_for_br_opp(self, step_wrappers):
""" Anticipatory; greedy BR + AVG """
SeatActorBase.act_mixed(
step_wrappers=step_wrappers,
br_learner=self.br_learner,
owner=self.owner,
avg_learner=self.avg_learner,
current_policy_tags=self._current_policy_tags_OPP_BR,
random_prob=0)
def act_for_avg_opp(self, step_wrappers):
"""
Purely random because that's how it should be for correct reach
"""
SeatActorBase.act_mixed(
step_wrappers=step_wrappers,
br_learner=self.br_learner,
owner=self.owner,
avg_learner=self.avg_learner,
current_policy_tags=self._current_policy_tags_O_AVG,
explore=True)
def act_for_br_trav(self, step_wrappers):
# Act
SeatActorBase.act_eps_greedy(step_wrappers=step_wrappers, br_learner=self.br_learner, owner=self.owner,
random_prob=self._constant_eps)
# Add to memories
for sw in step_wrappers:
e_i = sw.env_idx
self._br_memory_savers[e_i].add_experience(obs_t_before_acted=sw.obs,
a_selected_t=sw.action,
legal_actions_list_t=sw.legal_actions_list)
def update_if_terminal_for_br(self, step_wrappers, is_traverser=False):
for sw in step_wrappers:
if sw.TERMINAL:
if is_traverser:
self._br_memory_savers[sw.env_idx].add_to_buf(
reward_p=sw.term_rew_all[self.owner],
terminal_obs=sw.term_obs,
)
self._br_memory_savers[sw.env_idx].reset(
range_idx=sw.range_idxs[self.owner])
self._current_policy_tags_T_BR[
sw.env_idx] = SeatActorBase.pick_training_policy(
br_prob=self.sampler.antic)
else:
self._current_policy_tags_OPP_BR[
sw.env_idx] = SeatActorBase.pick_training_policy(
br_prob=self.sampler.antic)
def act_for_br_trav(self, step_wrappers):
# Act
SeatActorBase.act_mixed(
step_wrappers=step_wrappers,
br_learner=self.br_learner,
owner=self.owner,
avg_learner=self.avg_learner,
current_policy_tags=self._current_policy_tags_T_BR,
explore=True)
# Add to memories
for sw in step_wrappers:
e_i = sw.env_idx
self._br_memory_savers[e_i].add_experience(
obs_t_before_acted=sw.obs,
a_selected_t=sw.action,
legal_actions_list_t=sw.legal_actions_list)
def init(self, sws_br, sws_avg, nfsp_iter):
self._current_policy_tags_T_BR = np.empty(
shape=self._t_prof.n_steps_br_per_iter_per_la, dtype=np.int32)
self._current_policy_tags_OPP_BR = np.empty(
shape=self._t_prof.n_steps_br_per_iter_per_la, dtype=np.int32)
for sw in sws_br:
self._current_policy_tags_OPP_BR[
sw.env_idx] = SeatActorBase.pick_training_policy(
br_prob=self.sampler.antic)
self._current_policy_tags_T_BR[
sw.env_idx] = SeatActorBase.pick_training_policy(
br_prob=self.sampler.antic)
self._br_memory_savers[sw.env_idx].reset(
range_idx=sw.range_idxs[self.owner])
for sw in sws_avg:
self._avg_memory_savers[sw.env_idx].reset(
range_idx=sw.range_idxs[self.owner],
sample_weight=nfsp_iter if self._t_prof.linear else 1)
def update_if_terminal_for_avg(self,
step_wrappers,
nfsp_iter,
is_traverser=False):
for sw in step_wrappers:
if sw.TERMINAL:
if is_traverser:
self._avg_memory_savers[sw.env_idx].reset(
range_idx=sw.range_idxs[self.owner],
sample_weight=nfsp_iter if self._t_prof.linear else 1)
else:
self._current_policy_tags_O_AVG[
sw.env_idx] = SeatActorBase.pick_training_policy(
br_prob=self.sampler.antic)
def update_if_terminal(self, step_wrappers, nfsp_iter):
for sw in step_wrappers:
if sw.TERMINAL:
self._br_memory_savers[sw.env_idx].add_to_buf(
reward_p=sw.term_rew_all[self.seat_id],
terminal_obs=sw.term_obs,
)
self._br_memory_savers[sw.env_idx].reset(
range_idx=sw.range_idxs[self.seat_id])
self._avg_buf_savers[sw.env_idx].reset(
range_idx=sw.range_idxs[self.seat_id],
sample_weight=nfsp_iter if self._t_prof.linear else 1)
self._current_policy_tags[
sw.env_idx] = SeatActorBase.pick_training_policy(
br_prob=self.sampler.antic)
def act_for_avg_trav(self, step_wrappers):
""" BR greedy """
with torch.no_grad():
if len(step_wrappers) > 0:
actions, _ = SeatActorBase.choose_a_br(step_wrappers=step_wrappers, owner=self.owner,
br_learner=self.br_learner, random_prob=0)
for a, sw in zip(actions, step_wrappers):
a = a.item()
sw.action = a
sw.action_was_random = False
self._avg_memory_savers[sw.env_idx].add_step(pub_obs=sw.obs,
a=a,
legal_actions_mask=rl_util.get_legal_action_mask_np(
n_actions=self._env_bldr.N_ACTIONS,
legal_actions_list=sw.legal_actions_list)
)