def __init__(self, t_prof, seat_id, chief_handle):
self.ddqn_args = t_prof.module_args["ddqn"]
self.avg_args = t_prof.module_args["avg"]
super().__init__(t_prof=t_prof, chief_handle=chief_handle)
self.seat_id = seat_id
self.global_iter_id = 0
self.eps = self.ddqn_args.eps_start
self.antic = self._t_prof.antic_start
self.q_net = DuelingQNet(q_args=self.ddqn_args.q_args,
env_bldr=self._env_bldr,
device=self._device)
self.avg_net = AvrgStrategyNet(
avrg_net_args=self.avg_args.avg_net_args,
env_bldr=self._env_bldr,
device=self._device)
self.br_optim = rl_util.str_to_optim_cls(self.ddqn_args.optim_str)(
self.q_net.parameters(), lr=self.ddqn_args.lr)
self.avg_optim = rl_util.str_to_optim_cls(self.avg_args.optim_str)(
self.avg_net.parameters(), lr=self.avg_args.lr)
self.eps_exp = self._ray.remote(
self._chief_handle.create_experiment,
t_prof.name + ": epsilon Plyr" + str(seat_id))
self.antic_exp = self._ray.remote(
self._chief_handle.create_experiment,
t_prof.name + ": anticipatory Plyr" + str(seat_id))
self._log_eps()
self._log_antic()
def __init__(self, owner, env_bldr, avrg_training_args, device):
super().__init__(net=AvrgStrategyNet(
avrg_net_args=avrg_training_args.avrg_net_args,
env_bldr=env_bldr,
device=device),
env_bldr=env_bldr,
args=avrg_training_args,
owner=owner,
device=device)
self._all_range_idxs = torch.arange(self._env_bldr.rules.RANGE_SIZE,
device=self.device,
dtype=torch.long)
def _get_new_avrg_net(self):
return AvrgStrategyNet(avrg_net_args=self._avrg_args.avrg_net_args,
env_bldr=self._env_bldr,
device=self._device)