def get_action(self, hist_graph, curr_graph, tag2unit_dict):
assert isinstance(
curr_graph,
dgl.DGLGraph), "get action is designed to work on a single graph!"
num_time_steps = hist_graph.batch_size
hist_node_feature = hist_graph.ndata.pop('node_feature')
curr_node_feature = curr_graph.ndata.pop('node_feature')
maximum_num_enemy = get_largest_number_of_enemy_nodes([curr_graph])
nn_actions, info_dict = self.brain.get_action(
num_time_steps, hist_graph, hist_node_feature, curr_graph,
curr_node_feature, maximum_num_enemy)
ally_tags = info_dict['ally_tags']
enemy_tags = info_dict['enemy_tags']
sc2_actions = nn_action_to_sc2_action(nn_actions=nn_actions,
ally_tags=ally_tags,
enemy_tags=enemy_tags,
tag2unit_dict=tag2unit_dict)
hist_graph.ndata['node_feature'] = hist_node_feature
curr_graph.ndata['node_feature'] = curr_node_feature
return nn_actions, sc2_actions, info_dict
# print("=========={} th iter ============".format(i))
cur_state = env.observe()
# for tag, unit in cur_state['tag2unit_dict'].items():
# print("TAG: {} | NAME: {} | Health: {} | POS: {} ".format(tag, unit.name, unit.health, unit.position))
g = cur_state['g']
node_feature = g.ndata.pop('node_feature')
num_enemy = get_largest_number_of_enemy_nodes([g])
nn_actions, info_dict = ac_module.get_action(g, node_feature, num_enemy)
tag2unit_dict = cur_state['tag2unit_dict']
ally_tags = info_dict['ally_tags']
enemy_tags = info_dict['enemy_tags']
sc2_actions = nn_action_to_sc2_action(nn_actions=nn_actions,
ally_tags=ally_tags,
enemy_tags=enemy_tags,
tag2unit_dict=tag2unit_dict)
loss = ac_module(g, node_feature, num_enemy)
next_state, reward, done = env.step(action=sc2_actions)
if done:
done_cnt += 1
if done_cnt >= 10:
break
i += 1