2nd/10_Wiki/Topics/Domain_Programming/AI_and_ML/Micro-management.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-micro-management	Micro-management (RTS)	10_Wiki/Topics	verified	self	Micro RTS Micro Unit Micro StarCraft Micro	none	B	0.85	applied	game-design rts ai alphastar reinforcement-learning starcraft esports		2026-05-10	pending	language framework python pysc2|sc2|gymnasium

Micro-management (RTS)

한 줄: RTS에서 개별 유닛을 정밀 컨트롤(이동·공격·스킬)해 전투력을 극대화하는 행위 — APM·반응 속도·상황 인지의 종합. AlphaStar가 인간 상위 0.2%를 돌파한 RL 벤치마크.

핵심

• Macro vs Micro: macro = 경제·생산·확장, micro = 유닛 컨트롤. 프로는 둘 다.
• APM: Actions Per Minute. 인간 프로 200-400, AlphaStar 비공식 cap 280 (인간 수준 제한).
• 핵심 기술: focus fire, kiting (stutter step), spreading (스플래시 회피), flanking, target priority, 시거리 활용.
• AI 벤치마크: SC2 (DeepMind PySC2/AlphaStar), Dota 2 (OpenAI Five), Honor of Kings (JueWu) — long-horizon, partial observation, multi-agent의 종합 시험대.
• 현대 RL: self-play + league training (AlphaStar), PPO/IMPALA, large-scale distributed.

결정 기준

목표	접근
RTS AI 학습 환경	PySC2 (SC2), microRTS (가벼움), Stratagus
단순 micro 봇	스크립트 + finite state machine
강화학습	PPO + LSTM/transformer policy
멀티 유닛 협동	MARL (QMIX, MAPPO)
모방 학습 시작	Replay 데이터 → behavior cloning
대규모 학습	League / population-based training

💻 패턴

PySC2 minimal agent

from pysc2.agents import base_agent
from pysc2.lib import actions, features

class MicroAgent(base_agent.BaseAgent):
    def step(self, obs):
        super().step(obs)
        marines = [u for u in obs.observation.feature_units
                   if u.unit_type == features.PlayerRelative.SELF and u.alliance == 1]
        if not marines: return actions.FUNCTIONS.no_op()
        target = pick_lowest_hp_enemy(obs)
        if target is not None:
            return actions.FUNCTIONS.Attack_screen("now", [target.x, target.y])
        return actions.FUNCTIONS.no_op()

Kiting (stutter step) 의사 코드

def stutter_step(unit, enemy):
    if unit.weapon_ready:
        return Action.attack(enemy)
    if dist(unit, enemy) < unit.range and enemy.range > unit.range:
        return Action.move(away_from(enemy))
    return Action.move(toward(enemy))   # close gap

Focus fire (target priority)

def focus_target(my_units, enemies):
    # threat = damage * speed / hp_remaining
    return min(enemies, key=lambda e: e.hp / max(e.threat, 1))

PPO with PySC2 (sketch)

import gymnasium as gym
from stable_baselines3 import PPO
env = gym.make("SC2MoveToBeacon-v0")  # 단순 micro env
model = PPO("MultiInputPolicy", env, n_steps=2048, batch_size=64,
            learning_rate=3e-4, gamma=0.99, verbose=1)
model.learn(total_timesteps=2_000_000)

League training 개념 (AlphaStar)

# Main agents + Main exploiters + League exploiters; PFSP matchmaking
# Prioritized Fictitious Self-Play: 약한 상대일수록 자주 매칭하지 않음

Replay-based behavior cloning

# 1) replay parse → (state, action) pairs
# 2) supervised cross-entropy on actions
# 3) RL fine-tune from BC checkpoint (jumpstart RL exploration)

🔗 Graph

• 상위: Reinforcement-Learning
• 관련: PPO

🤖 LLM 활용

• 게임 디자인 토론(LLM brainstorm): 새 유닛 micro skill 설계, 카운터 분석.
• 코드 어시스트: PySC2 보일러플레이트, replay parser.
• LLM 단독으로 RTS 플레이는 비현실적 (실시간·연속 액션 부적합) — RL과 결합.

❌ 안티패턴

• APM만 추구 — 의미 없는 클릭 (spam). EPM (effective APM)이 중요.
• 모든 유닛 attack-move — 마이크로 포기. 부대 분리·focus fire·kiting 조합.
• RL 환경에서 sparse reward만 — 학습 안 됨. shaping (kill, hp 차이) + curriculum.
• single-agent로 multi-unit RTS 학습 — 정책 collapse. 적절한 abstraction (군 단위) 또는 MARL.
• AlphaStar 재현 시도 with 작은 GPU — 수만 TPU-day 필요. microRTS·SMAC 같은 축소 환경부터.

🧪 검증 / 중복

• 중복 후보 없음.
• 검증: 표준 task — MoveToBeacon, CollectMineralShards, DefeatRoaches (PySC2 mini-games)·SMAC 시나리오 win-rate.

🕓 Changelog

• 2026-05-08 | Phase 1 — 자동 시드.
• 2026-05-10 | Manual cleanup — PySC2/PPO/league 패턴, kiting 의사코드, RL 안티패턴 정리.