# CollectMineralShards backup 저는 이번 튜토리얼을 통해, AlphaStar논문에서의 Agent Network에서 적용된 AutoRegressive model을 모방해, 화면 RGB와 action은 두 개의 마린과 mineralshards들을 entity로 모두 넣을 생각입니다.(화면 RGB만으로 충분할것으로 보이지만) 기본적으로 제가 생각한 필요 정보들은 다음과 같습니다. * 화면의 정보 * observation.screen\_features.player\_relative 를 사용할 예정입니다.(마린은 1, Mineralshards 은 3) * 화면의 마린과 Mineralshards(dynamic length) * x좌표 * observation.feature\_units.x * y좌표 * observation.feature\_units.y * alliance * observation.feature\_units.alliance * (Mineralshards는 3, marine은 1) 그렇다면, state를 구하는 function을 만들어보겠습니다. ```python def get_state(obs): screen = obs.observation.feature_screen.player_relative units_info = obs.observation.feature_units units_info = [[unit.x,unit.y,unit.alliance] for unit in units_info] return np.expand_dims(np.array(screen),0), np.expand_dims(np.array(units_info),0) ``` 그렇다면 action은 어떻게 정의할 수 있을까요? 저는 위에서 말했듯이 AutoRegressive 모델을 만들기 위해 다음과 같이 정의하여보았습니다. * Action * unit selection * mineralshards * marine * target unit selection * marine * mineralshards * none * target point * x,y 이들을 통합해 만든 스크립트입니다. Brain은 현재 마구잡이 random으로 action하도록 만들어진 상태입니다. ```python MOVE_SCREEN = 331 NOT_QUEUED = [0] SELF = features.PlayerRelative.SELF NEUTRAL = features.PlayerRelative.NEUTRAL MINERALSHARDS = 1680 MARINES = 48 class Brain: def __init__(self): self.action_lst = [] def network(self, obs,screen_info,units_info): ## network를 잠시 function형태로 쓰겠습니다. units_info = obs.observation.feature_units units_info = [[unit.alliance,unit.x,unit.y] for unit in units_info] #모든 selection풀기 지금 deselect 기능 못쓰므로 잘못눌렀다면 no_op()해주어야함 unit = random.choice(units_info) if unit[0] == NEUTRAL: return actions.FUNCTIONS.no_op(), actions.FUNCTIONS.no_op(), actions.FUNCTIONS.no_op() else: action_1 = actions.FUNCTIONS.select_point("select",(unit[1],unit[2])) if random.randint(0,1) == 0: # target unit이 잡혔을 시 target_unit = random.choice(units_info) action_2 = actions.FunctionCall(MOVE_SCREEN,[NOT_QUEUED,[target_unit[1],target_unit[2]]]) action_3 = actions.FUNCTIONS.no_op() else: # target unit없을시 action_2 = actions.FUNCTIONS.no_op x = random.randint(0,SCREEN_SIZE - 1) y = random.randint(0,SCREEN_SIZE - 1) action_3 = actions.FunctionCall(MOVE_SCREEN,[NOT_QUEUED,[x,y]]) return action_1,action_2,action_3 def action(self,obs,screen_info,units_info): if len(self.action_lst) == 0: [self.action_lst.append(x) for x in self.network(obs,screen_info,units_info)] return self.action_lst.pop(0) class Agent(base_agent.BaseAgent): def __init__(self): super(Agent, self).__init__() self.brain = Brain() def get_state(self,obs): screen = obs.observation.feature_screen.player_relative units_info = obs.observation.feature_units units_info = [[unit.x,unit.y,unit.alliance] for unit in units_info] return np.expand_dims(np.array(screen),0), np.expand_dims(np.array(units_info),0) # .transpose(1, 0) def step(self, obs): super(Agent, self).step(obs) screen_info, units_info = self.get_state(obs) action = self.brain.action(obs,screen_info,units_info) #self.brain.train() return action ``` 다음은 reward를 받아와보겠습니다. 위의 스크립트로 본다면 98번째 줄에서 reward를 확인해야합니다. ```python try: with sc2_env.SC2Env(map_name=MAPNAME, players=players, \ agent_interface_format=interface, \ step_mul=APM, game_steps_per_episode=UNLIMIT, \ visualize=VISUALIZE, realtime=REALTIME) as env: agent.setup(env.observation_spec(), env.action_spec()) timestep = env.reset() agent.reset() reward = 0 while True: step_actions = [agent.step(timestep[0])] if timestep[0].last(): break timestep = env.step(step_actions) reward -= timestep[0].observation.player.minerals except KeyboardInterrupt: pass ``` 그렇다면 전체 코드는 다음과 같습니다. ```python from pysc2.env import sc2_env from pysc2.lib import features from pysc2.agents import base_agent import numpy as np from pysc2.lib import actions from absl import app import random MAPNAME = 'CollectMineralShards' APM = 300 APM = int(APM / 18.75) UNLIMIT = 0 VISUALIZE = True REALTIME = True SCREEN_SIZE = 84 MINIMAP_SIZE = 64 MOVE_SCREEN = 331 NOT_QUEUED = [0] SELF = features.PlayerRelative.SELF NEUTRAL = features.PlayerRelative.NEUTRAL MINERALSHARDS = 1680 MARINES = 48 players = [sc2_env.Agent(sc2_env.Race.terran)] interface = features.AgentInterfaceFormat( \ feature_dimensions=features.Dimensions( \ screen=SCREEN_SIZE, minimap=MINIMAP_SIZE), use_feature_units=True) class Brain: def __init__(self): self.action_lst = [] def network(self, obs,screen_info,units_info): ## network를 잠시 function형태로 쓰겠습니다. units_info = obs.observation.feature_units units_info = [[unit.alliance,unit.x,unit.y] for unit in units_info] #모든 selection풀기 지금 deselect 기능 못쓰므로 잘못눌렀다면 no_op()해주어야함 unit = random.choice(units_info) if random.randint(0,1) == 0: #아무것도 안하는 행위(이미 마린들이 목표를향해 움직인다면 아무것도안하고 있어도 괜찮음) return actions.FUNCTIONS.no_op(), actions.FUNCTIONS.no_op(), actions.FUNCTIONS.no_op() elif unit[0] == NEUTRAL: return actions.FUNCTIONS.no_op(), actions.FUNCTIONS.no_op(), actions.FUNCTIONS.no_op() else: action_1 = actions.FUNCTIONS.select_point("select",(unit[1],unit[2])) if random.randint(0,1) == 0: # target unit이 잡혔을 시 target_unit = random.choice(units_info) action_2 = actions.FunctionCall(MOVE_SCREEN,[NOT_QUEUED,[target_unit[1],target_unit[2]]]) action_3 = actions.FUNCTIONS.no_op() else: # target unit없을시 action_2 = actions.FUNCTIONS.no_op() x = random.randint(0,SCREEN_SIZE - 1) y = random.randint(0,SCREEN_SIZE - 1) action_3 = actions.FunctionCall(MOVE_SCREEN,[NOT_QUEUED,[x,y]]) return action_1,action_2,action_3 def action(self,obs,screen_info,units_info): if len(self.action_lst) == 0: [self.action_lst.append(x) for x in self.network(obs,screen_info,units_info)] return self.action_lst.pop(0) class Agent(base_agent.BaseAgent): def __init__(self): super(Agent, self).__init__() self.brain = Brain() def get_state(self,obs): screen = obs.observation.feature_screen.player_relative units_info = obs.observation.feature_units units_info = [[unit.x,unit.y,unit.alliance] for unit in units_info] return np.expand_dims(np.array(screen),0), np.expand_dims(np.array(units_info),0) # .transpose(1, 0) def step(self, obs): super(Agent, self).step(obs) screen_info, units_info = self.get_state(obs) action = self.brain.action(obs,screen_info,units_info) #self.brain.train() return action def main(args): agent = Agent() try: with sc2_env.SC2Env(map_name=MAPNAME, players=players, \ agent_interface_format=interface, \ step_mul=APM, game_steps_per_episode=UNLIMIT, \ visualize=VISUALIZE, realtime=REALTIME) as env: agent.setup(env.observation_spec(), env.action_spec()) timestep = env.reset() agent.reset() while True: reward = - timestep[0].observation.player.minerals step_actions = [agent.step(timestep[0])] if timestep[0].last(): break timestep = env.step(step_actions) reward += timestep[0].observation.player.minerals except KeyboardInterrupt: pass app.run(main) ``` 완료입니다! 이후 테스트하고싶은 것은 CollectMineralsAndGas 맵에서 MultiObjective로 미네랄과 가스의 weights 를 주어 weights에 따라 미네랄과 가스의 채취량을 조절할 수 있는가에 대한 테스트를 개인적으로 진행할 예정입니다. 번외로, units로 screen에 안보이더라도 확인가능한지 --- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://kilmya1.gitbook.io/startcraft2-rl-agent-tutorial/solve-mini-game/mini-game/collectmineralshards-backup.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.