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 _get_new_avrg_optim(self):
opt = rl_util.str_to_optim_cls(self._avrg_args.optim_str)(
self._avrg_net.parameters(), lr=self._avrg_args.lr)
scheduler = lr_scheduler.ReduceLROnPlateau(
optimizer=opt,
threshold=0.0001,
factor=0.5,
patience=self._avrg_args.lr_patience,
min_lr=0.00002)
return opt, scheduler
def _get_new_optim(self, p_id):
return rl_util.str_to_optim_cls(self._args.ddqn_args.optim_str)(
self._nets[p_id].parameters(), lr=self._args.ddqn_args.lr)
def _get_new_baseline_optim(self):
opt = rl_util.str_to_optim_cls(self._baseline_args.optim_str)(
self._baseline_net.parameters(), lr=self._baseline_args.lr)
return opt