def _get_nl_leduc_tree(env_args=None):
if env_args is None:
env_args = DiscretizedNLLeduc.ARGS_CLS(
n_seats=2,
starting_stack_sizes_list=[1000, 1000],
bet_sizes_list_as_frac_of_pot=[1.0])
env_bldr = HistoryEnvBuilder(env_cls=DiscretizedNLLeduc, env_args=env_args)
_tree = PublicTree(
env_bldr=env_bldr,
stack_size=env_args.starting_stack_sizes_list,
stop_at_street=None,
)
_tree.build_tree()
for p in range(env_bldr.N_SEATS):
_tree.fill_uniform_random()
_tree.compute_ev()
_tree.export_to_file()
print("Tree with stack size", _tree.stack_size, "has", _tree.n_nodes,
"nodes out of which", _tree.n_nonterm, "are non-terminal.")
print(np.mean(_tree.root.exploitability) * env_bldr.env_cls.EV_NORMALIZER)
return _tree
def _evaluate_avg_strats(self):
expl_totals = []
for t_idx in range(len(self._trees)):
METRIC = self._env_bldrs[t_idx].env_cls.WIN_METRIC
eval_tree = PublicTree(
env_bldr=self._env_bldrs[t_idx],
stack_size=self._env_args[t_idx].starting_stack_sizes_list,
stop_at_street=None,
is_debugging=False,
)
eval_tree.build_tree()
def _fill(_node_eval, _node_train):
if _node_eval.p_id_acting_next != eval_tree.CHANCE_ID and (
not _node_eval.is_terminal):
_node_eval.strategy = np.copy(
_node_train.data["avg_strat"])
assert np.allclose(np.sum(_node_eval.strategy, axis=1),
1,
atol=0.0001)
for c_eval, c_train in zip(_node_eval.children,
_node_train.children):
_fill(_node_eval=c_eval, _node_train=c_train)
# sets up some stuff; we overwrite strategy afterwards
eval_tree.fill_uniform_random()
# fill with strat
_fill(_node_eval=eval_tree.root,
_node_train=self._trees[t_idx].root)
eval_tree.update_reach_probs()
# compute EVs
eval_tree.compute_ev()
eval_tree.export_to_file(name=self._name + "_Avg_" +
str(self._iter_counter))
# log
expl_p = [
float(eval_tree.root.exploitability[p]) *
self._env_bldrs[t_idx].env_cls.EV_NORMALIZER
for p in range(eval_tree.n_seats)
]
expl_total = sum(expl_p) / eval_tree.n_seats
expl_totals.append(expl_total)
self._chief_handle.add_scalar(self._exps_avg_total[t_idx],
"Evaluation/" + METRIC,
self._iter_counter, expl_total)
expl_total_averaged = sum(expl_totals) / float(len(expl_totals))
self._chief_handle.add_scalar(self._exp_all_averaged_avg_total,
"Evaluation/" + METRIC,
self._iter_counter, expl_total_averaged)
return expl_total_averaged