def guess_action_cal_reward(state_path, output_path):
state_file = open(state_path, "r")
output = open(output_path, 'w')
lines = state_file.readlines()
state_logs = []
prev_state = None
for line in lines:
cur_state = StateUtil.parse_state_log(line)
if cur_state.tick == StateUtil.TICK_PER_STATE:
print("clear")
prev_state = None
elif prev_state is not None and prev_state.tick >= cur_state.tick:
print ("clear")
prev_state = None
if prev_state is not None:
state_logs.append(prev_state)
prev_state = cur_state
if prev_state is not None:
state_logs.append(prev_state)
# 猜测玩家行为
for idx in range(1, len(state_logs)-1):
prev_state = state_logs[idx-1]
cur_state = state_logs[idx]
next_state = state_logs[idx+1]
if cur_state.tick >= 55044:
db = 1
hero = prev_state.get_hero("27")
line_index = 1
near_enemy_heroes = StateUtil.get_nearby_enemy_heros(prev_state, hero.hero_name,
StateUtil.LINE_MODEL_RADIUS)
near_enemy_units = StateUtil.get_nearby_enemy_units(prev_state, hero.hero_name, StateUtil.LINE_MODEL_RADIUS)
nearest_enemy_tower = StateUtil.get_nearest_enemy_tower(prev_state, hero.hero_name,
StateUtil.LINE_MODEL_RADIUS)
near_enemy_units_in_line = StateUtil.get_units_in_line(near_enemy_units, line_index)
nearest_enemy_tower_in_line = StateUtil.get_units_in_line([nearest_enemy_tower], line_index)
if len(near_enemy_heroes) != 0 or len(near_enemy_units_in_line) != 0 or len(
nearest_enemy_tower_in_line) != 0:
player_action = Replayer.guess_player_action(prev_state, cur_state, next_state, "27", "28")
action_str = StateUtil.build_command(player_action)
print('玩家行为分析：' + str(action_str) + ' tick:' + str(prev_state.tick) + ' prev_pos: ' +
hero.pos.to_string() + ', cur_pos: ' + cur_state.get_hero(hero.hero_name).pos.to_string())
prev_state.add_action(player_action)
# 测试计算奖励值
state_logs_with_reward = LineModel.update_rewards(state_logs)
for state_with_reward in state_logs_with_reward:
# 将结果记录到文件
state_encode = state_with_reward.encode()
state_json = JSON.dumps(state_encode)
output.write(strftime("%Y-%m-%d %H:%M:%S", gmtime()) + " -- " + state_json + "\n")
output.flush()
print(len(state_logs))
def upgrade_skills(self, state_info, hero_name):
# 如果有可以升级的技能，优先升级技能3
hero = state_info.get_hero(hero_name)
skills = StateUtil.get_skills_can_upgrade(hero)
if len(skills) > 0:
skillid = 3 if 3 in skills else skills[0]
update_cmd = CmdAction(hero.hero_name, CmdActionEnum.UPDATE,
skillid, None, None, None, None, None, None)
update_str = StateUtil.build_command(update_cmd)
return update_str
def guess_hero_actions(self, state_index, real_heros=None):
prev_state = self.state_cache[state_index - 1]
cur_state = self.state_cache[state_index]
next_state = self.state_cache[state_index + 1]
# 如果有必要的话，更新这一帧中真人玩家的行为信息
if real_heros is not None:
for hero_name in real_heros:
hero_action = Replayer.guess_player_action(prev_state, cur_state, next_state, hero_name, '28')
cur_state.add_action(hero_action)
action_str = StateUtil.build_command(hero_action)
print('玩家行为分析：' + str(action_str) + ' tick:' + str(cur_state.tick))
def add_other_hero_action(input_data, hero_info, action_cmd, debug=False):
# 如果是自己，则忽略
if hero_info.hero_name == action_cmd.hero_name:
return input_data
# 如果不是攻击类行为，忽略
if action_cmd.action != CmdActionEnum.ATTACK and action_cmd.action != CmdActionEnum.CAST:
return input_data
# 如果不是己方的动作，忽略
friends, opponents = TeamBattleUtil.get_friend_opponent_heros(
TeamBattleInput.HERO_LIST, hero_info.hero_name)
if action_cmd.hero_name != hero_info.hero_name and action_cmd.hero_name not in friends:
return input_data
# 更新输入数据
# 首先找到目标英雄ID，然后找到使用的技能ID
hero_index = friends.index(action_cmd.hero_name) + 1
tgt_hero_index = TeamBattleInput.TEAM_A.index(action_cmd.tgtid) if action_cmd.tgtid in TeamBattleInput.TEAM_A \
else TeamBattleInput.TEAM_B.index(action_cmd.tgtid)
change_index = hero_index * 89 + 15 + tgt_hero_index if action_cmd.action == CmdActionEnum.ATTACK \
else hero_index * 89 + 20 + (int(action_cmd.skillid) - 1) * 23 + 18 + tgt_hero_index
prev_value = input_data[change_index]
if prev_value != 0:
if debug:
print("add_other_hero_action", "must be something wrong",
"prev_value not zero")
input_data[change_index] = 1
debug_action_str = StateUtil.build_command(action_cmd)
if debug:
print("add_other_hero_action", "add_hero_info",
hero_info.hero_name, "hero_index", hero_index,
"tgt_hero_index", tgt_hero_index, "action_cmd.action",
action_cmd.action, "action_cmd_skill", action_cmd.skillid,
"change_index", change_index, "cmd", debug_action_str)
return input_data
def build_response(self, state_cache, state_index, hero_name):
action_strs=[]
restart = False
# 对于模型，分析当前帧的行为
if self.real_hero != hero_name:
state_info = state_cache[state_index]
prev_hero = state_cache[state_index-1].get_hero(hero_name) if len(state_cache) >= 2 is not None else None
# 如果有真实玩家，我们需要一些历史数据，所以分析3帧前的行为
elif len(state_cache) > 3:
state_info = state_cache[state_index-3]
next1_state_info = state_cache[state_index-2]
next2_state_info = state_cache[state_index-1]
next3_state_info = state_cache[state_index]
else:
return action_strs, False
# 决定是否购买道具
buy_action = EquipUtil.buy_equip(state_info, hero_name)
if buy_action is not None:
buy_str = StateUtil.build_command(buy_action)
action_strs.append(buy_str)
# 如果有可以升级的技能，优先升级技能3
hero = state_info.get_hero(hero_name)
skills = StateUtil.get_skills_can_upgrade(hero)
if len(skills) > 0:
skillid = 3 if 3 in skills else skills[0]
update_cmd = CmdAction(hero.hero_name, CmdActionEnum.UPDATE, skillid, None, None, None, None, None, None)
update_str = StateUtil.build_command(update_cmd)
action_strs.append(update_str)
# 回城相关逻辑
# 如果在回城中且没有被打断则继续回城，什么也不用返回
if prev_hero is not None:
if hero.hero_name in self.hero_strategy and self.hero_strategy[hero.hero_name] == ActionEnum.town_ing \
and prev_hero.hp <= hero.hp \
and not StateUtil.if_hero_at_basement(hero):
if not hero.skills[6].canuse:
print(self.battle_id, hero.hero_name, '回城中，继续回城')
return action_strs, False
else:
print(self.battle_id, hero.hero_name, '回城失败')
town_action = CmdAction(hero.hero_name, CmdActionEnum.CAST, 6, hero.hero_name, None, None, None,
None, None)
action_str = StateUtil.build_command(town_action)
action_strs.append(action_str)
return action_strs, False
if hero.hp <= 0:
self.hero_strategy[hero.hero_name] = None
return action_strs, False
# # 补血逻辑
# if prev_hero is not None and hero.hero_name in self.hero_strategy and self.hero_strategy[
# hero.hero_name] == ActionEnum.hp_restore:
# if StateUtil.cal_distance2(prev_hero.pos, hero.pos) < 100:
# print(self.battle_id, hero_name, '到达补血点', '血量增长', hero.hp - prev_hero.hp)
# del self.hero_strategy[hero_name]
# if hero == self.model1_hero:
# self.model1_hp_restore = time.time()
# else:
# self.model2_hp_restore = time.time()
# 撤退逻辑
# TODO 甚至可以使用移动技能移动
if prev_hero is not None and hero.hero_name in self.hero_strategy and self.hero_strategy[hero.hero_name] == ActionEnum.retreat_to_town:
if StateUtil.cal_distance2(prev_hero.pos, hero.pos) < 100:
print(self.battle_id, hero_name, '开始回城')
self.hero_strategy[hero.hero_name] = ActionEnum.town_ing
town_action = CmdAction(hero.hero_name, CmdActionEnum.CAST, 6, hero.hero_name, None, None, None,
None, None)
action_str = StateUtil.build_command(town_action)
action_strs.append(action_str)
else:
print(self.battle_id, hero_name, '还在撤退中', StateUtil.cal_distance2(prev_hero.pos, hero.pos))
return action_strs, False
# 如果击杀了对方英雄，扫清附近小兵之后则启动撤退回城逻辑
if prev_hero is not None:
if hero.hero_name in self.hero_strategy and self.hero_strategy[hero.hero_name] == ActionEnum.town_ing and prev_hero.hp <= hero.hp \
and not StateUtil.if_hero_at_basement(hero):
if not hero.skills[6].canuse:
return action_strs, False
else:
town_action = CmdAction(hero.hero_name, CmdActionEnum.CAST, 6, hero.hero_name, None, None, None,
None, None)
action_str = StateUtil.build_command(town_action)
action_strs.append(action_str)
if hero.hp <= 0:
self.hero_strategy[hero.hero_name] = None
return action_strs, False
# 检查周围状况
near_enemy_heroes = StateUtil.get_nearby_enemy_heros(state_info, hero.hero_name, StateUtil.LINE_MODEL_RADIUS)
near_enemy_units = StateUtil.get_nearby_enemy_units(state_info, hero.hero_name, StateUtil.LINE_MODEL_RADIUS)
nearest_enemy_tower = StateUtil.get_nearest_enemy_tower(state_info, hero.hero_name,
StateUtil.LINE_MODEL_RADIUS + 3)
nearest_friend_units = StateUtil.get_nearby_friend_units(state_info, hero.hero_name, StateUtil.LINE_MODEL_RADIUS)
line_index = 1
near_enemy_units_in_line = StateUtil.get_units_in_line(near_enemy_units, line_index)
nearest_enemy_tower_in_line = StateUtil.get_units_in_line([nearest_enemy_tower], line_index)
# 如果击杀对面英雄就回城补血。整体逻辑为，周围没有兵的情况下启动撤退逻辑，到达撤退地点之后启动回城。补满血之后再跟兵出来
# 处在泉水之中的时候设置策略层为吃线
if len(near_enemy_units_in_line) == 0 and len(near_enemy_heroes) == 0:
if (hero_name == self.model1_hero and self.model2_just_dead == 1 and not StateUtil.if_hero_at_basement(hero)) \
or (hero_name == self.model2_hero and self.model1_just_dead == 1 and not StateUtil.if_hero_at_basement(hero)):
if hero.hp / float(hero.maxhp) > 0.8:
if hero_name == self.model1_hero:
self.model2_just_dead = 0
else:
self.model1_just_dead = 0
else:
print(self.battle_id, hero_name, '选择撤退')
self.hero_strategy[hero_name] = ActionEnum.retreat_to_town
retreat_pos = StateUtil.get_retreat_pos(state_info, hero, line_index=1)
action = CmdAction(hero_name, CmdActionEnum.MOVE, None, None, retreat_pos, None, None, -1, None)
action_str = StateUtil.build_command(action)
action_strs.append(action_str)
if hero_name == self.model1_hero:
self.model2_just_dead = 0
else:
self.model1_just_dead = 0
return action_strs, False
if StateUtil.if_hero_at_basement(hero):
if hero_name == self.model1_hero:
self.model2_just_dead = 0
else:
self.model1_just_dead = 0
if hero.hp < hero.maxhp:
if hero_name in self.hero_strategy:
del self.hero_strategy[hero_name]
return action_strs, False
# # 残血并且周围没有敌人的情况下，可以去塔后吃加血
# if hero.hp / float(hero.maxhp) < 0.9 and hero not in self.hero_strategy:
# print('补血条件', self.battle_id, hero_name, time.time(), self.model1_hp_restore, self.model2_hp_restore)
# if hero == self.model1_hero and time.time() - self.model1_hp_restore > LineTrainerPPO.HP_RESTORE_GAP:
# print(self.battle_id, hero_name, '选择加血')
# self.hero_strategy[hero_name] = ActionEnum.hp_restore
# elif hero == self.model2_hero and time.time() - self.model2_hp_restore > LineTrainerPPO.HP_RESTORE_GAP:
# print(self.battle_id, hero_name, '选择加血')
# self.hero_strategy[hero_name] = ActionEnum.hp_restore
#
# if self.hero_strategy[hero_name] == ActionEnum.hp_restore:
# restore_pos = StateUtil.get_hp_restore_place(state_info, hero)
# action = CmdAction(hero_name, CmdActionEnum.MOVE, None, None, restore_pos, None, None, -1, None)
# action_str = StateUtil.build_command(action)
# action_strs.append(action_str)
# return action_strs, False
# 开始根据策略决定当前的行动
# 对线情况下，首先拿到兵线，朝最前方的兵线移动
# 如果周围有危险（敌方单位）则启动对线模型
# 如果周围有小兵或者塔，需要他们都是在指定线上的小兵或者塔
if (len(near_enemy_units_in_line) == 0 and len(nearest_enemy_tower_in_line) == 0 and (
len(near_enemy_heroes) == 0 or
StateUtil.if_in_line(hero, line_index, 4000) == -1)
) or\
(len(nearest_friend_units) == 0 and len(near_enemy_units_in_line) == 0 and
len(near_enemy_heroes) == 0 and len(nearest_enemy_tower_in_line) == 1):
# 跟兵线或者跟塔，优先跟塔
self.hero_strategy[hero.hero_name] = ActionEnum.line_1
# print("策略层：因为附近没有指定兵线的敌人所以开始吃线 " + hero.hero_name)
front_soldier = StateUtil.get_frontest_soldier_in_line(state_info, line_index, hero.team)
first_tower = StateUtil.get_first_tower(state_info, hero)
if front_soldier is None or (hero.team == 0 and first_tower.pos.x > front_soldier.pos.x) or (hero.team == 1 and first_tower.pos.x < front_soldier.pos.x):
# 跟塔，如果塔在前面的话
follow_tower_pos = StateUtil.get_tower_behind(first_tower, hero, line_index=1)
move_action = CmdAction(hero.hero_name, CmdActionEnum.MOVE, None, None, follow_tower_pos, None, None,
None, None)
action_str = StateUtil.build_command(move_action)
action_strs.append(action_str)
else:
# 得到最前方的兵线位置
move_action = CmdAction(hero.hero_name, CmdActionEnum.MOVE, None, None, front_soldier.pos, None, None,
None, None)
action_str = StateUtil.build_command(move_action)
action_strs.append(action_str)
else:
if self.real_hero != hero_name:
# 使用模型进行决策
# print("使用对线模型决定英雄%s的行动" % hero.hero_name)
self.hero_strategy[hero.hero_name] = ActionEnum.line_model
# 目前对线只涉及到两名英雄
rival_hero = '28' if hero.hero_name == '27' else '27'
action, explorer_ratio, action_ratios = self.get_action(state_info, hero_name, rival_hero)
# 考虑使用固定策略
# 如果决定使用策略，会连续n条行为全都采用策略（比如确保对方残血时候连续攻击的情况）
# 如果策略返回为空则表示策略中断
if self.policy_ratio > 0 and (
0 < self.cur_policy_act_idx_map[hero_name] < self.policy_continue_acts
or random.uniform(0, 1) <= self.policy_ratio
):
policy_action = LineTrainerPolicy.choose_action(state_info, action_ratios, hero_name, rival_hero,
near_enemy_units, nearest_friend_units)
if policy_action is not None:
policy_action.vpred = action.vpred
action = policy_action
self.cur_policy_act_idx_map[hero_name] += 1
print("英雄 " + hero_name + " 使用策略，策略行为计数 idx " + str(self.cur_policy_act_idx_map[hero_name]))
if self.cur_policy_act_idx_map[hero_name] >= self.policy_continue_acts:
self.cur_policy_act_idx_map[hero_name] = 0
else:
# 策略中断，清零
if self.cur_policy_act_idx_map[hero_name] > 0:
print("英雄 " + hero_name + " 策略中断，清零")
self.cur_policy_act_idx_map[hero_name] = 0
action_str = StateUtil.build_command(action)
action_strs.append(action_str)
# 如果是要求英雄施法回城，更新英雄状态，这里涉及到后续多帧是否等待回城结束
if action.action == CmdActionEnum.CAST and int(action.skillid) == 6:
print("英雄%s释放了回城" % hero_name)
self.hero_strategy[hero.hero_name] = ActionEnum.town_ing
# 如果是选择了撤退，进行特殊标记，会影响到后续的行为
if action.action == CmdActionEnum.RETREAT:
print("英雄%s释放了撤退，撤退点为%s" % (hero_name, action.tgtpos.to_string()))
self.hero_strategy[hero.hero_name] = ActionEnum.retreat
self.retreat_pos = action.tgtpos
# 如果批量训练结束了，这时候需要清空未使用的训练集，然后重启游戏
if action.action == CmdActionEnum.RESTART:
restart = True
else:
# 保存action信息到状态帧中
state_info.add_action(action)
else:
# 还是需要模型来计算出一个vpred
rival_hero = '28' if hero.hero_name == '27' else '27'
action, explorer_ratio, action_ratios = self.get_action(state_info, hero_name, rival_hero)
# 推测玩家的行为
guess_action = Replayer.guess_player_action(state_info, next1_state_info, next2_state_info,
next3_state_info, hero_name, rival_hero)
guess_action.vpred = action.vpred
action_str = StateUtil.build_command(guess_action)
action_str['tick'] = state_info.tick
print('猜测玩家行为为：' + JSON.dumps(action_str))
# 保存action信息到状态帧中
state_info.add_action(guess_action)
return action_strs, restart
def buy_equip(self, state_info, hero_name):
# 决定是否购买道具
buy_action = EquipUtil.buy_equip(state_info, hero_name)
if buy_action is not None:
buy_str = StateUtil.build_command(buy_action)
return buy_str
def build_response(self, raw_state_str):
self.save_raw_log(raw_state_str)
prev_state_info = self.state_cache[-1] if len(
self.state_cache) > 0 else None
response_strs = []
# 解析客户端发送的请求
obj = JSON.loads(raw_state_str)
raw_state_info = StateInfo.decode(obj)
# 重开时候会有以下报文 {"wldstatic":{"ID":9051},"wldruntime":{"State":0}}
if raw_state_info.tick == -1:
return {"ID": raw_state_info.battleid, "tick": -1}
if raw_state_info.tick <= StateUtil.TICK_PER_STATE and (
prev_state_info is None
or prev_state_info.tick > raw_state_info.tick):
print("clear")
prev_state_info = None
self.state_cache = []
self.battle_started = -1
self.battle_heroes_cache = []
self.dead_heroes = []
self.dead_heroes_cache = []
self.data_inputs = []
self.rebooting = False
elif prev_state_info is None and raw_state_info.tick > StateUtil.TICK_PER_STATE:
# 不是开始帧的话直接返回重启游戏
# 还有偶然情况下首帧没有tick（即-1）的情况，这种情况下只能重启本场战斗
print("battle_id", self.battle_id, "tick", raw_state_info.tick,
'不是开始帧的话直接返回重启游戏', raw_state_info.tick)
action_strs = [
StateUtil.build_action_command('27', 'RESTART', None)
]
rsp_obj = {
"ID": raw_state_info.battleid,
"tick": raw_state_info.tick,
"cmd": action_strs
}
rsp_str = JSON.dumps(rsp_obj)
return rsp_str
state_info = StateUtil.update_state_log(prev_state_info,
raw_state_info)
hero = state_info.get_hero("27")
if hero is None or hero.hp is None:
# 偶然情况处理，如果找不到英雄，直接重开
print("battle_id", self.battle_id, "tick", state_info.tick,
'不是开始帧的话直接返回重启游戏', raw_state_info.tick)
action_strs = [
StateUtil.build_action_command('27', 'RESTART', None)
]
rsp_obj = {
"ID": raw_state_info.battleid,
"tick": raw_state_info.tick,
"cmd": action_strs
}
rsp_str = JSON.dumps(rsp_obj)
return rsp_str
# 战斗前准备工作
if len(self.state_cache) == 0:
# 第一帧的时候，添加金钱和等级
for hero in self.heros:
add_gold_cmd = CmdAction(hero, CmdActionEnum.ADDGOLD, None,
None, None, None, None, None, None)
add_gold_cmd.gold = 3000
add_gold_str = StateUtil.build_command(add_gold_cmd)
response_strs.append(add_gold_str)
add_lv_cmd = CmdAction(hero, CmdActionEnum.ADDLV, None, None,
None, None, None, None, None)
add_lv_cmd.lv = 9
add_lv_str = StateUtil.build_command(add_lv_cmd)
response_strs.append(add_lv_str)
elif len(self.state_cache) > 1:
# 第二帧时候开始，升级技能，购买装备，这个操作可能会持续好几帧
for hero in self.heros:
upgrade_cmd = self.upgrade_skills(state_info, hero)
if upgrade_cmd is not None:
response_strs.append(upgrade_cmd)
buy_cmd = self.buy_equip(state_info, hero)
if buy_cmd is not None:
response_strs.append(buy_cmd)
for hero in self.heros:
# 判断是否英雄死亡
if prev_state_info is not None:
dead = StateUtil.if_hero_dead(prev_state_info, state_info,
hero)
if dead == 1 and hero not in self.dead_heroes:
print("battle_id", self.battle_id, "tick", state_info.tick,
"英雄死亡", hero, "tick", state_info.tick)
self.dead_heroes.append(hero)
# 首先要求所有英雄站到团战圈内，然后开始模型计算，这时候所有的行动都有模型来决定
# 需要过滤掉无效的行动，同时屏蔽会离开战斗圈的移动
#TODO 开始团战后，如果有偶尔的技能移动会离开圈，则拉回来
# 这里会排除掉死亡的英雄，他们不需要再加入团战
# 团战范围在收缩
battle_range = self.cal_battle_range(
len(self.state_cache) - self.battle_started)
heroes_in_range, heroes_out_range = TeamBattleTrainer.all_in_battle_range(
state_info, self.heros, self.dead_heroes, battle_range)
# 存活英雄
battle_heros = list(heroes_in_range)
battle_heros.extend(heroes_out_range)
# 缓存参战情况和死亡情况，用于后续训练
self.battle_heroes_cache.append(battle_heros)
self.dead_heroes_cache.append(list(self.dead_heroes))
if state_info.tick >= 142560:
debuginfo = True
# 团战还没有开始，有英雄还在圈外
if len(heroes_out_range) > 0:
if self.battle_started > -1:
print('battle_id', self.battle_id, "战斗已经开始，但是为什么还有英雄在团战圈外",
','.join(heroes_out_range), "battle_range", battle_range)
# 移动到两个开始战斗地点附近
# 如果是团战开始之后，移动到团战中心点
for hero in heroes_out_range:
start_point_x = randint(0, 8000)
start_point_z = TeamBattleTrainer.BATTLE_CIRCLE_RADIUS_BATTLE_START * 1000 if self.battle_started == -1 else 0
start_point_z += randint(-4000, 4000)
if TeamBattleUtil.get_hero_team(hero) == 0:
start_point_z *= -1
start_point_z += TeamBattleTrainer.BATTLE_POINT_Z
tgt_pos = PosStateInfo(start_point_x, 0, start_point_z)
move_action = CmdAction(hero, CmdActionEnum.MOVE, None, None,
tgt_pos, None, None, None, None)
mov_cmd_str = StateUtil.build_command(move_action)
response_strs.append(mov_cmd_str)
# 团战已经开始
elif not self.rebooting:
if self.battle_started == -1:
self.battle_started = len(self.state_cache)
# 对特殊情况。比如德古拉使用大招hp会变1，修改帧状态
state_info, _ = TeamBattlePolicy.modify_status_4_draculas_invincible(
state_info, self.state_cache)
# action_cmds, input_list, model_upgrade = self.get_model_actions(state_info, heroes_in_range)
# 跟队伍，每个队伍得到行为
team_a, team_b = TeamBattleUtil.get_teams(heroes_in_range)
team_actions_a, input_list_a, model_upgrade_a = self.get_model_actions_team(
state_info, team_a, heroes_in_range)
team_actions_b, input_list_b, model_upgrade_b = self.get_model_actions_team(
state_info, team_b, heroes_in_range)
# 如果模型已经开战，重启战斗
if (model_upgrade_a or model_upgrade_b
) and self.battle_started < len(self.state_cache) + 1:
print("battle_id", self.battle_id, "因为模型升级，重启战斗",
self.battle_started, len(self.state_cache))
action_strs = [
StateUtil.build_action_command('27', 'RESTART', None)
]
rsp_obj = {
"ID": raw_state_info.battleid,
"tick": raw_state_info.tick,
"cmd": action_strs
}
rsp_str = JSON.dumps(rsp_obj)
return rsp_str
data_input_map = {}
for action_cmd, data_input in zip(team_actions_a + team_actions_b,
input_list_a + input_list_b):
action_str = StateUtil.build_command(action_cmd)
response_strs.append(action_str)
state_info.add_action(action_cmd)
data_input_map[action_cmd.hero_name] = data_input
# 缓存所有的模型输入，用于后续训练
self.data_inputs.append(data_input_map)
# 添加记录到缓存中
self.state_cache.append(state_info)
# 将模型行为加入训练缓存，同时计算奖励值
# 注意：因为奖励值需要看后续状态，所以这个计算会有延迟
last_x_index = 2
if self.battle_started > -1 and len(self.data_inputs) >= last_x_index:
if self.rebooting:
# 测试发现重启指令发出之后，可能下一帧还没开始重启战斗，这种情况下抛弃训练
print("battle_id", self.battle_id, "tick", state_info.tick,
"warn", "要求重启战斗，但是还在收到后续帧状态, 继续重启")
# 重启游戏
response_strs = [
StateUtil.build_action_command('27', 'RESTART', None)
]
else:
state_index = len(self.state_cache) - last_x_index
win, win_team, left_heroes = self.remember_replay_heroes(
-last_x_index, state_index, battle_range)
# 团战结束条件
# 首先战至最后一人
# all_in_team = TeamBattleUtil.all_in_one_team(heroes_in_range)
# if self.battle_started:
# if len(self.dead_heroes) >= 9 or (len(self.dead_heroes) >= 5 and all_in_team > -1):
if win == 1:
# 重启游戏
print('battle_id', self.battle_id, "重启游戏", "剩余人员",
','.join(left_heroes))
response_strs = [
StateUtil.build_action_command('27', 'RESTART', None)
]
self.rebooting = True
# battle_heros = self.search_team_battle(state_info)
# if len(battle_heros) > 0:
# print("team battle heros", ';'.join(battle_heros))
#
# heros_need_model = []
# for hero in self.heros:
# # 判断是否英雄死亡
# if prev_state_info is not None:
# dead = StateUtil.if_hero_dead(prev_state_info, state_info, hero)
# if dead == 1 and hero not in self.dead_heroes:
# self.dead_heroes.append(hero)
#
# # 复活的英雄不要再去参团
# if hero in self.dead_heroes:
# continue
#
# # near_enemy_heroes = StateUtil.get_nearby_enemy_heros(state_info, hero, TeamBattleTrainer.MODEL_RANGE)
# if hero not in battle_heros:
# # 移动到团战点附近，添加部分随机
# rdm_delta_x = randint(0, 1000)
# rdm_delta_z = randint(0, 1000)
# tgt_pos = PosStateInfo(TeamBattleTrainer.BATTLE_POINT_X + rdm_delta_x, 0, TeamBattleTrainer.BATTLE_POINT_Z + rdm_delta_z)
# move_action = CmdAction(hero, CmdActionEnum.MOVE, None, None, tgt_pos, None, None, None, None)
# mov_cmd_str = StateUtil.build_command(move_action)
# response_strs.append(mov_cmd_str)
# else:
# # 启动模型决策
# heros_need_model.append(hero)
#
# if len(heros_need_model) > 0:
# action_cmds = self.get_model_actions(state_info, heros_need_model)
# for action_cmd in action_cmds:
# action_str = StateUtil.build_command(action_cmd)
# response_strs.append(action_str)
# state_info.add_action(action_cmd)
#TODO 记录模型输出，用于后续训练
# 返回结果给游戏端
rsp_obj = {
"ID": state_info.battleid,
"tick": state_info.tick,
"cmd": response_strs
}
rsp_str = JSON.dumps(rsp_obj)
print('battle_id', self.battle_id, 'response', rsp_str)
return rsp_str
def get_model_actions(self, state_info, heros, debug=False):
# 第一个人先选，然后第二个人，一直往后，后面的人会在参数中添加上之前人的行为
# TODO 同时可以变成按照模型给出maxq大小来决定谁先选
# 这样的好处是所有人选择的行为就是最后执行的行为
# 暂时为随机英雄先选
random_heros = list(heros)
shuffle(random_heros)
# 得到当前团战范围，因为会收缩
battle_range = self.cal_battle_range(
len(self.state_cache) - self.battle_started)
action_cmds = []
input_list = []
model_upgrade = False
for hero in random_heros:
hero_info = state_info.get_hero(hero)
data_input = TeamBattleInput.gen_input(state_info, hero)
data_input = np.array(data_input)
# 对于之前的英雄行为，加入输入
for prev_action in action_cmds:
data_input = TeamBattleInput.add_other_hero_action(
data_input, hero_info, prev_action, debug)
action_list, explor_value, vpreds, clear_cache = self.model_util.get_action_list(
self.battle_id, hero, data_input)
action_str = ' '.join(
str("%.4f" % float(act)) for act in action_list)
if debug:
print("battle_id", self.battle_id, "tick", state_info.tick,
"hero", hero, "model action list", action_str)
unaval_list = TeamBattleTrainer.list_unaval_actions(
action_list, state_info, hero, heros, battle_range)
unaval_list_str = ' '.join(
str("%.4f" % float(act)) for act in unaval_list)
if debug:
print("battle_id", self.battle_id, "tick", state_info.tick,
"hero", hero, "model remove_unaval_actions",
unaval_list_str)
friends, opponents = TeamBattleUtil.get_friend_opponent_heros(
heros, hero)
action_cmd, max_q, selected = TeamBattleTrainer.get_action_cmd(
action_list, unaval_list, state_info, hero, friends, opponents)
if debug:
print("battle_id", self.battle_id, "tick", state_info.tick,
"hero", hero, "model get_action",
StateUtil.build_command(action_cmd), "max_q", max_q,
"selected", selected)
# 如果模型升级了，需要清空所有缓存用作训练的行为，并且重启游戏
if clear_cache:
print('battle_id', self.battle_id, '模型升级，清空训练缓存')
for hero_name in self.heros:
self.model_caches[hero_name].clear_cache()
model_upgrade = True
action_cmds.append(action_cmd)
input_list.append(data_input)
return action_cmds, input_list, model_upgrade
def get_model_actions_team(self,
state_info,
team,
battle_heroes,
debug=False):
# 第一个人先选，然后第二个人，一直往后，后面的人会在参数中添加上之前人的行为
# 同时可以变成按照模型给出maxq大小来决定谁先选
# 这样的好处是所有人选择的行为就是最后执行的行为
# 暂时为随机英雄先选
# first_hero = heroes[0]
# 得到当前团战范围，因为会收缩
battle_range = self.cal_battle_range(
len(self.state_cache) - self.battle_started)
# 首先得到当前情况下每个英雄的基础输入集和所有无效的选择
hero_input_map = {}
hero_unavail_list_map = {}
for hero in team:
data_input = TeamBattleInput.gen_input(state_info, hero,
battle_heroes)
data_input = np.array(data_input)
hero_input_map[hero] = data_input
unaval_list = TeamBattleTrainer.list_unaval_actions(
self.act_size, state_info, hero, battle_heroes, battle_range)
unaval_list_str = ' '.join(
str("%.4f" % float(act)) for act in unaval_list)
hero_unavail_list_map[hero] = unaval_list
if debug:
print("battle_id", self.battle_id, "tick", state_info.tick,
"hero", hero, "model remove_unaval_actions",
unaval_list_str)
# 得到每个英雄的推荐行为
hero_recommend_list_map = {}
for hero in team:
friends, opponents = TeamBattleUtil.get_friend_opponent_heros(
battle_heroes, hero)
hero_info = state_info.get_hero(hero)
recommend_list = TeamBattlePolicy.select_action_by_strategy(
state_info, hero_info, friends, opponents)
hero_recommend_list_map[hero] = recommend_list
# 开始挑选英雄行为，每次根据剩余英雄的最优选择，根据Q大小来排序
action_cmds = []
input_list = []
left_heroes = list(team)
model_upgrade = False
while len(left_heroes) > 0:
cur_max_q = -1
chosen_hero = left_heroes[0]
chosen_action_list = None
for hero in left_heroes:
# 对于之前的英雄行为，加入输入
hero_info = state_info.get_hero(hero)
data_input = hero_input_map[hero]
for prev_action in action_cmds:
data_input = TeamBattleInput.add_other_hero_action(
data_input, hero_info, prev_action, debug)
unaval_list = hero_unavail_list_map[hero]
recommend_list = hero_recommend_list_map[hero]
action_list, explor_value, vpreds, clear_cache = self.model_util.get_action_list(
self.battle_id, hero, data_input)
action_str = ' '.join(
str("%.4f" % float(act)) for act in action_list)
max_q = TeamBattleTrainer.get_max_q(action_list, unaval_list,
recommend_list)
if debug:
print("battle_id", self.battle_id, "tick", state_info.tick,
"本轮行为候选", "hero", hero, "max_q", max_q,
"model action list", action_str)
# 允许等于是为了支持max_q等于-1的情况
if max_q >= cur_max_q:
cur_max_q = max_q
chosen_hero = hero
chosen_action_list = action_list
# 如果模型升级了，需要清空所有缓存用作训练的行为，并且重启游戏
if clear_cache:
print('battle_id', self.battle_id, '模型升级，清空训练缓存')
for hero_name in self.heros:
self.model_caches[hero_name].clear_cache()
model_upgrade = True
# 使用最大q的英雄的行为
unaval_list = hero_unavail_list_map[chosen_hero]
recommend_list = hero_recommend_list_map[hero]
friends, opponents = TeamBattleUtil.get_friend_opponent_heros(
battle_heroes, chosen_hero)
action_cmd, max_q, selected = TeamBattleTrainer.get_action_cmd(
chosen_action_list, unaval_list, recommend_list, state_info,
chosen_hero, friends, opponents)
if debug:
print("battle_id", self.battle_id, "tick", state_info.tick,
"hero", chosen_hero, "model get_action",
StateUtil.build_command(action_cmd), "max_q", max_q,
"selected", selected)
# 更新各个状态集
action_cmds.append(action_cmd)
input_list.append(data_input)
left_heroes.remove(chosen_hero)
return action_cmds, input_list, model_upgrade
def build_response(self, state_info, prev_state_info, line_model, hero_names=None):
battle_id = state_info.battleid
tick = state_info.tick
if tick >= 139062:
db = 1
action_strs=[]
if hero_names is None:
hero_names = [hero.hero_name for hero in state_info.heros]
for hero_name in hero_names:
hero = state_info.get_hero(hero_name)
prev_hero = prev_state_info.get_hero(hero.hero_name) if prev_state_info is not None else None
# 检查是否重启游戏
# 线上第一个塔被摧毁
# 如果有可以升级的技能，优先升级技能3
skills = StateUtil.get_skills_can_upgrade(hero)
if len(skills) > 0:
skillid = 3 if 3 in skills else skills[0]
update_cmd = CmdAction(hero.hero_name, CmdActionEnum.UPDATE, skillid, None, None, None, None, None, None)
update_str = StateUtil.build_command(update_cmd)
action_strs.append(update_str)
# 检查周围状况
near_enemy_heroes = StateUtil.get_nearby_enemy_heros(state_info, hero.hero_name, StateUtil.LINE_MODEL_RADIUS)
near_enemy_units = StateUtil.get_nearby_enemy_units(state_info, hero.hero_name, StateUtil.LINE_MODEL_RADIUS)
nearest_enemy_tower = StateUtil.get_nearest_enemy_tower(state_info, hero.hero_name, StateUtil.LINE_MODEL_RADIUS + 3)
# 回城相关逻辑
# 如果在回城中且没有被打断则继续回城，什么也不用返回
if prev_hero is not None:
if self.hero_strategy[hero.hero_name] == ActionEnum.town_ing and prev_hero.hp <= hero.hp \
and not StateUtil.if_hero_at_basement(hero):
if not hero.skills[6].canuse:
print('回城中，继续回城')
continue
else:
print('回城失败')
if hero.hp <= 0:
self.hero_strategy[hero.hero_name] = None
continue
# 处在少血状态是，且周围没有地方单位的情况下选择回城
# if len(near_enemy_heroes) == 0 and len(near_enemy_units) == 0 and nearest_enemy_tower is None:
# if hero.hp/float(hero.maxhp) < LineTrainer.TOWN_HP_THRESHOLD:
# print('策略层：回城')
# # 检查英雄当前状态，如果在回城但是上一帧中受到了伤害，则将状态设置为正在回城，开始回城
# if self.hero_strategy[hero.hero_name] == ActionEnum.town_ing:
# if prev_hero.hp > hero.hp:
# town_action = CmdAction(hero.hero_name, CmdActionEnum.CAST, 6, hero.hero_name, None, None, None, None, None)
# action_str = StateUtil.build_command(town_action)
# action_strs.append(action_str)
# # 检查英雄当前状态，如果不在回城，则将状态设置为正在回城，开始回城
# elif self.hero_strategy[hero.hero_name] != ActionEnum.town_ing:
# self.hero_strategy[hero.hero_name] = ActionEnum.town_ing
# town_action = CmdAction(hero.hero_name, CmdActionEnum.CAST, 6, hero.hero_name, None, None, None, None, None)
# action_str = StateUtil.build_command(town_action)
# action_strs.append(action_str)
#
# # 无论上面怎么操作，玩家下面的动作应该都是在回城中，所以跳过其它的操作
# continue
# 处在泉水之中的时候设置策略层为吃线
if StateUtil.if_hero_at_basement(hero):
if hero.hp < hero.maxhp:
continue
# 撤退逻辑
# TODO 甚至可以使用移动技能移动
if hero.hero_name in self.hero_strategy and self.hero_strategy[hero.hero_name] == ActionEnum.retreat:
dist = StateUtil.cal_distance(hero.pos, self.retreat_pos)
if dist <= 2:
print('到达撤退点附近')
self.hero_strategy[hero.hero_name] = None
elif prev_hero is not None and prev_hero.pos.to_string() == hero.pos.to_string():
print('英雄卡住了，取消撤退')
self.hero_strategy[hero.hero_name] = None
else:
print('仍然在撤退 ' + str(dist))
continue
# 开始根据策略决定当前的行动
# 对线情况下，首先拿到兵线，朝最前方的兵线移动
# 如果周围有危险（敌方单位）则启动对线模型
# 如果周围有小兵或者塔，需要他们都是在指定线上的小兵或者塔
line_index = 1
near_enemy_units_in_line = StateUtil.get_units_in_line(near_enemy_units, line_index)
nearest_enemy_tower_in_line = StateUtil.get_units_in_line([nearest_enemy_tower], line_index)
if len(near_enemy_units_in_line) == 0 and len(nearest_enemy_tower_in_line) == 0 and (len(near_enemy_heroes) == 0 or
StateUtil.if_in_line(hero, line_index, 4000) == -1):
self.hero_strategy[hero.hero_name] = ActionEnum.line_1
# print("策略层：因为附近没有指定兵线的敌人所以开始吃线 " + hero.hero_name)
# 跟兵线
front_soldier = StateUtil.get_frontest_soldier_in_line(state_info, line_index, hero.team)
if front_soldier is None:
action_str = StateUtil.build_action_command(hero.hero_name, 'HOLD', {})
action_strs.append(action_str)
else:
# 得到最前方的兵线位置
move_action = CmdAction(hero.hero_name, CmdActionEnum.MOVE, None, None, front_soldier.pos, None, None, None, None)
action_str = StateUtil.build_command(move_action)
action_strs.append(action_str)
else:
# 使用模型进行决策
# print("使用对线模型决定英雄%s的行动" % hero.hero_name)
self.hero_strategy[hero.hero_name] = ActionEnum.line_model
enemies = []
enemies.extend((hero.hero_name for hero in near_enemy_heroes))
enemies.extend((unit.unit_name for unit in near_enemy_units))
if nearest_enemy_tower is not None:
enemies.append(nearest_enemy_tower.unit_name)
# print('对线模型决策，因为周围有敌人 ' + ' ,'.join(enemies))
# 目前对线只涉及到两名英雄
rival_hero = '28' if hero.hero_name == '27' else '27'
action = line_model.get_action(prev_state_info, state_info, hero.hero_name, rival_hero)
action_str = StateUtil.build_command(action)
action_strs.append(action_str)
# 如果是要求英雄施法回城，更新英雄状态，这里涉及到后续多帧是否等待回城结束
if action.action == CmdActionEnum.CAST and int(action.skillid) == 6:
print("英雄%s释放了回城" % hero_name)
self.hero_strategy[hero.hero_name] = ActionEnum.town_ing
# 如果是选择了撤退，进行特殊标记，会影响到后续的行为
if action.action == CmdActionEnum.RETREAT:
print("英雄%s释放了撤退，撤退点为%s" % (hero_name, action.tgtpos.to_string()))
self.hero_strategy[hero.hero_name] = ActionEnum.retreat
self.retreat_pos = action.tgtpos
# 保存action信息到状态帧中
state_info.add_action(action)
return action_strs