ゲームバランス調整 ── LLM プレイヤーで自動プレイテスト

from dataclasses import dataclass, asdictfrom typing import Literal
@dataclassclass GameState:    player_hp: int    player_max_hp: int    player_mp: int    enemies: list[dict]  # [{"name": str, "hp": int}]    items: list[str]    location: str    turn: int
@dataclassclass Action:    type: Literal["attack", "magic", "item", "flee"]    target: str | None = None    item_name: str | None = None    magic_name: str | None = None
def legal_actions(state: GameState) -> list[Action]:    actions = []    for enemy in state.enemies:        actions.append(Action("attack", target=enemy["name"]))    if state.player_mp >= 5:        actions.append(Action("magic", magic_name="fire"))    for item in state.items:        actions.append(Action("item", item_name=item))    actions.append(Action("flee"))    return actions

from anthropic import Anthropicimport jsonimport os
client = Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])
PERSONAS = {    "speedrunner": "最短でクリア、回復は最小限、危険でも前進。",    "collector": "全アイテム回収、隠し要素探索、戦闘は丁寧に。",    "cautious": "HP 半分以下なら必ず回復、危険な敵は逃げる。",    "berserker": "常に攻撃魔法、回復軽視、最大火力。",    "newbie": "あまり考えずランダム寄り、たまに操作ミスで弱い行動。",}
def llm_choose_action(state: GameState, persona: str) -> Action:    actions = legal_actions(state)    actions_text = "\n".join(        f"{i}: {asdict(a)}" for i, a in enumerate(actions)    )
    prompt = f"""あなたは {persona} 型のプレイヤー: {PERSONAS[persona]}
現在の状態:{json.dumps(asdict(state), ensure_ascii=False, indent=2)}
選択可能な行動:{actions_text}
理由を一言と、選んだ番号を JSON で返してください:{{"reason": "...", "choice": 0}}"""
    response = client.messages.create(        model="claude-haiku-4-5-20251001",        max_tokens=100,        messages=[{"role": "user", "content": prompt}],    )    text = response.content[0].text
    # JSON 抽出 (失敗時はランダム)    try:        decision = json.loads(text[text.index("{"):text.rindex("}")+1])        return actions[decision["choice"]]    except Exception:        import random        return random.choice(actions)

import statisticsfrom collections import defaultdict
def simulate_run(stage_id: str, persona: str) -> dict:    """1 試行のシミュレーション (結果を辞書で返す)"""    state = load_stage(stage_id)  # ステージ初期状態    turn = 0    death_locations = []    while not is_clear(state) and not is_dead(state) and turn < 200:        action = llm_choose_action(state, persona)        state = apply_action(state, action)        if is_dead(state):            death_locations.append(state.location)        turn += 1    return {        "stage_id": stage_id,        "persona": persona,        "cleared": is_clear(state),        "turns": turn,        "remaining_hp": state.player_hp,        "death_locations": death_locations,    }
# 全ステージ × 全ペルソナで 20 試行results = []for stage_id in ["stage_01", "stage_02", "stage_03"]:    for persona in PERSONAS:        for trial in range(20):            results.append(simulate_run(stage_id, persona))
# 集計summary = defaultdict(lambda: {"clear": 0, "total": 0, "turns": []})for r in results:    key = (r["stage_id"], r["persona"])    summary[key]["total"] += 1    if r["cleared"]:        summary[key]["clear"] += 1        summary[key]["turns"].append(r["turns"])
print(f"{'Stage':12} {'Persona':12} {'ClearRate':10} {'AvgTurns':10}")for (sid, p), s in summary.items():    rate = s["clear"] / s["total"] * 100    avg = statistics.mean(s["turns"]) if s["turns"] else 0    print(f"{sid:12} {p:12} {rate:>8.1f}% {avg:>8.1f}")

import matplotlib.pyplot as pltimport numpy as npfrom collections import Counter
# 全試行から死亡地点を集計all_deaths = Counter()for r in results:    for loc in r["death_locations"]:        all_deaths[loc] += 1
# マップを 2D グリッドに見立ててヒートマップ# (実際はマップ座標 → グリッドインデックスの変換が要る)grid = np.zeros((20, 30))LOC_TO_XY = {    "entry": (0, 0), "corridor_a": (5, 10), "trap_room": (10, 15),    # ...}for loc, count in all_deaths.items():    if loc in LOC_TO_XY:        y, x = LOC_TO_XY[loc]        grid[y, x] = count
plt.figure(figsize=(10, 6))plt.imshow(grid, cmap='hot', interpolation='nearest')plt.colorbar(label='Death count')plt.title('Death locations heatmap')plt.show()

# MCTS の rollout で 「完全ランダム」 でなく# 「LLM が選んだ妥当な手」 でロールアウトすると精度向上# (AlphaGo の policy network の代わりに LLM を使う)
def llm_rollout(state):    """MCTS の simulation 段階で LLM に方策を提供"""    while not is_terminal(state):        # 簡易 LLM (Haiku) で手を選ばせる        action = llm_choose_action(state, persona="optimal")        state = apply_action(state, action)    return evaluate(state)  # 勝敗 / スコア