iter9-numfix-b: 模拟档位重构——普通档 v3=完美+失误注入·完美档修上界（用户拍板「普通档≠新手画像」）
- 依据：真人实证推翻 v2 普通档——零加成真人首通三关踩 23 雷 vs 新手画像 L2 条件通关率 10.6% （尺子比真人弱 20 倍）；完美档选路机械取最小编号+商店零购=不是元游戏上界（真人两店全购） - 普通档 v3：完美策略（subset 解算/标旗/卡死刷新）+受控失误注入——标雷误挖 NORMAL_MINE_SLIP_P 0.22（目标踩雷期望=完美+2/板，用户规格）+安全挖跳过 NORMAL_SKIP_DIG_P 0.05（每板少挖几个）； 失误流=独立策略 RNG（局种子派生，逐局可回放；核心七流零消费零扰动） - 完美档上界修正：选路机械最小编号→目标价值评分（rest 伤重/shop 买得起/reward 占位/boss 缓冲， 同分 tie-break 沿 route 方向——verifier 双窗轴保持）；商店零购→贪心买齐（GP 局内死钱） - 夹具保护：v3 行为只挂 run_adventure 测量面（st.adv 旗标）——run_level 单关机制夹具保持旧语义 （机械选路+novice 画像+商店零购），smoke 断言零扰动 - novice_action 留库不删（历史读数对账面）；轻闸=冒烟 596/0 全绿（夹具零漂移实证）
This commit is contained in:
DKChan committed
1 parent
2301db5540
commit
a2f72f9d9e
1 file changed
+130
-12
+130
-12
@@ -18,12 +18,27 @@
|
||||
# 大奖励立即确认 continue_to_next_level、adventure_finished 置 result/summary；外层每关重取
|
||||
# runner 实例经 _drive_level 复用单关驱动）。run_level 驱动段抽成 _drive_level——签名与
|
||||
# 返回结构零变化，行为逐位不变。
|
||||
# 2026-10-01 模拟档位重构（用户拍板「普通档≠新手画像」——普通档 v2 被真人实证推翻：零加成真人
|
||||
# 首通三关踩 23 雷 vs 新手画像 L2 条件通关率 10.6%，尺子比真人弱 20 倍）：
|
||||
# ①普通档 v3 = 完美策略 + 受控失误注入（标雷误挖 NORMAL_MINE_SLIP_P / 安全挖跳过
|
||||
# NORMAL_SKIP_DIG_P；失误流走独立策略 RNG——种子派生自局种子，逐局确定性/可回放保持，
|
||||
# 核心七流零消费零扰动）；v2「新手级单步推理+盲挖兜底」口径退役（novice_action 留库不删——
|
||||
# 历史读数对账面）。
|
||||
# ②完美档修上界：选路从机械最小编号改目标价值评分（rest 伤重优先/shop 有钱优先/boss 缓冲），
|
||||
# 商店从零购离开改贪心买齐（GP 局内死钱——不购物=玩的不是同一个游戏，真人实证两店全购）。
|
||||
# ③route 轴保持：评分同分 tie-break 仍按 min/max 编号方向——verifier 双窗轴不塌。
|
||||
# 运行前提：db 已 load_tables；run_level/run_adventure 均须全新 RngService（同种子 ⇒ 全程逐步可回放）。
|
||||
# 实现注：信号处理体全部为静态方法 + bind 传状态——GDScript 多行 lambda 内嵌 match/控制流的解析不可靠。
|
||||
extends RefCounted
|
||||
|
||||
const AutoPlayer := preload("res://smoke/auto_player.gd") # O4 收口：自动玩家策略单一实现
|
||||
|
||||
# ── 普通档 v3 失误注入参数（2026-10-01 用户规格：踩雷期望=完美+2/板、每板少挖几个正确格）──
|
||||
# 标雷→误挖转化率：每板已证雷 ~9-14 颗 × p ≈ 2-3 次踩雷（目标完美+2/板±0.3——首测读数后校准此值）
|
||||
const NORMAL_MINE_SLIP_P := 0.22
|
||||
# 安全挖跳过率：每板安全格 ~70-90 × p ≈ 3-5 次漏挖（信息与推进双损失）
|
||||
const NORMAL_SKIP_DIG_P := 0.05
|
||||
|
||||
|
||||
static func run_level(db, seed_val: int, level_index: int, atk_pct: int, hp_pct: int, collect_rows: bool = false, logger = null, policy: String = "normal", route: String = "min", template_pick: int = -1) -> Dictionary:
|
||||
var own_session := logger == null
|
||||
@@ -42,6 +57,7 @@ static func run_level(db, seed_val: int, level_index: int, atk_pct: int, hp_pct:
|
||||
"boss_depth": -1, "hp_at_boss": -1, "rows": [], "timeout": false,
|
||||
"log_path": "", "policy": policy, "refreshes": 0, "mine_hits": 0,
|
||||
"visited_ids": [], "choices": 0, "last_battle_enemies": [],
|
||||
"prng": _policy_rng(seed_val), "mine_slips": 0, "skip_digs": 0, "shop_buys": 0,
|
||||
}
|
||||
runner.step_row.connect(_on_step_row.bind(runner, st, collect_rows))
|
||||
runner.node_entered.connect(_on_node_entered.bind(runner, st, logger))
|
||||
@@ -80,7 +96,7 @@ static func _drive_level(runner, st: Dictionary, logger, policy: String, route:
|
||||
# 选路态（入口停点 / 兑付确认 / 战斗胜利后的零自动推进停点）——route 策略显式选下一跳
|
||||
if runner.node_index >= 0:
|
||||
st.choices = int(st.choices) + 1
|
||||
runner.pick_next_node(_route_pick(runner, route))
|
||||
runner.pick_next_node(_route_pick(runner, route, st))
|
||||
continue
|
||||
guard_battles -= 1
|
||||
var this_settlement = runner.current_battle()
|
||||
@@ -118,6 +134,8 @@ static func run_adventure(db, seed_val: int, atk_pct: int, hp_pct: int, collect_
|
||||
"visited_ids": [], "choices": 0, "last_battle_enemies": [],
|
||||
"summary": {}, "wallet_events": [], "encounter_max_hp_seq": [], "reward_events": [],
|
||||
"shop_gp_seq": [], # 商店房到达余额序（逐关 1 次——GP 跨关延续验收取数面）
|
||||
"prng": _policy_rng(seed_val), "mine_slips": 0, "skip_digs": 0, "shop_buys": 0,
|
||||
"adv": true, # v3 档位行为开关（选路评分/失误注入/商店购买只挂冒险测量面——单关夹具保持旧语义零扰动）
|
||||
|
||||
"level_templates": [], # 逐关图模板名（关卡节点图生成 v0.4 分层出数）
|
||||
}
|
||||
@@ -175,7 +193,8 @@ static func _on_adventure_node_entered(node: Dictionary, info: Dictionary, adr,
|
||||
r.continue_after_node()
|
||||
elif t == "shop":
|
||||
st.shop_gp_seq.append(int(info.get("gp", -1)))
|
||||
r.continue_after_node() # 商店房零购离开（自动路径直通 Boss——购买行为归冒烟专项/真人试玩）
|
||||
_autoplay_shop(r, st) # 贪心买齐（2026-10-01——旧零购离开退役：GP 局内死钱，真人实证两店全购）
|
||||
r.continue_after_node()
|
||||
if logger != null:
|
||||
logger.log_node(node, info, r.node_index)
|
||||
|
||||
@@ -195,6 +214,28 @@ static func _on_adventure_reward(reward: Dictionary, st: Dictionary, adr) -> voi
|
||||
st.result = "advance_failed"
|
||||
|
||||
|
||||
# 商店贪心买齐（2026-10-01 模拟档位重构）：GP 局内死钱——买得起就买（ affordable 集里挑最便宜的
|
||||
# 反复买到达不上任何商品为止，guard 防死循环）。确定性（零随机流）；购买出账/加算区入账全走
|
||||
# runner.buy_product 核心编排（与呈现层购买同一出账路径，wallet_changed 信号照常采集）。
|
||||
static func _autoplay_shop(r, st: Dictionary) -> void:
|
||||
var guard := 30
|
||||
while guard > 0:
|
||||
guard -= 1
|
||||
var affordable: Array = []
|
||||
for p in r.shop_products():
|
||||
if r.gp >= int(p.get("price", 1 << 30)):
|
||||
affordable.append(p)
|
||||
if affordable.is_empty():
|
||||
return
|
||||
var pick: Dictionary = affordable[0]
|
||||
for p in affordable:
|
||||
if int(p.get("price", 1 << 30)) < int(pick.get("price", 1 << 30)):
|
||||
pick = p
|
||||
if not r.buy_product(String(pick.get("product_id", ""))):
|
||||
return
|
||||
st.shop_buys = int(st.get("shop_buys", 0)) + 1
|
||||
|
||||
|
||||
# 冒险终局（completed/failed/abandoned——result 载 summary.result，外层驱动循环随之收敛）
|
||||
static func _on_adventure_finished(summary: Dictionary, st: Dictionary, logger) -> void:
|
||||
st.summary = summary.duplicate(true)
|
||||
@@ -203,14 +244,58 @@ static func _on_adventure_finished(summary: Dictionary, st: Dictionary, logger)
|
||||
logger.log_level_end(summary)
|
||||
|
||||
|
||||
# 路径轴策略（确定性——候选编号最小/最大；min 侧 = 岔路恒取小编号支）。
|
||||
# 编号取 runner.next_candidate_ids()（编号空间解析单一出口——单链下标 / 图 node_depth）
|
||||
static func _route_pick(runner, route: String) -> int:
|
||||
# 策略流 RNG（2026-10-01 模拟档位重构）：种子派生自局种子 × 黄金比率常数 + 偏移——与核心统一流
|
||||
# （rng_service 单盐派生）零共享零扰动；逐局确定性保持（同种子 ⇒ 同决策序列 ⇒ 同回放）
|
||||
static func _policy_rng(seed_val: int) -> RandomNumberGenerator:
|
||||
var prng := RandomNumberGenerator.new()
|
||||
prng.seed = seed_val * 2654435761 + 912931
|
||||
return prng
|
||||
|
||||
|
||||
# 路径轴选路（2026-10-01 目标价值评分制——机械最小编号退役）：候选按节点类型评分取最高，
|
||||
# 同分 tie-break 沿 route 方向（min=小编号/max=大编号——verifier 双窗轴保持）。
|
||||
# 评分语义（确定性，零随机流）：rest=伤重高值（缺口比例×100）/shop=买得起高值（GP 局内死钱，
|
||||
# 真人实证两店全购）/reward=占位低值（MC 夹具解锁空集=零兑付直通）/boss=缓冲低值（有店有_rest
|
||||
# 先补给再打）/battle=基线。编号取 runner.next_candidate_ids()（编号空间解析单一出口）。
|
||||
static func _route_pick(runner, route: String, st: Dictionary) -> int:
|
||||
var ids: Array = runner.next_candidate_ids()
|
||||
if ids.is_empty():
|
||||
return -1
|
||||
ids.sort()
|
||||
return ids[0] if route != "max" else ids[ids.size() - 1]
|
||||
if not bool(st.get("adv", false)): # 单关机制夹具：旧机械选路保持（smoke 断言零扰动）
|
||||
ids.sort()
|
||||
return int(ids[0]) if route != "max" else int(ids[ids.size() - 1])
|
||||
var nodes: Array = runner.next_candidates()
|
||||
var scored: Array = []
|
||||
for i in nodes.size():
|
||||
var node: Dictionary = nodes[i]
|
||||
var t := String(node.get("type", ""))
|
||||
var s := 30.0
|
||||
if t == "rest":
|
||||
var max_hp := float(runner.effective_max_hp())
|
||||
s = 5.0 + (1.0 - float(runner.hero_hp) / maxf(max_hp, 1.0)) * 100.0 # 满血 rest=5 低于 battle（白 rest 不去）；伤重可到 105
|
||||
elif t == "shop":
|
||||
s = 10.0
|
||||
var cheapest := 1 << 30
|
||||
for p in runner.shop_products():
|
||||
cheapest = mini(cheapest, int(p.get("price", 1 << 30)))
|
||||
if cheapest < (1 << 30) and runner.gp >= cheapest:
|
||||
s = 80.0
|
||||
elif t == "reward":
|
||||
s = 5.0
|
||||
elif t == "boss":
|
||||
s = 20.0
|
||||
scored.append({"id": int(ids[i]), "s": s})
|
||||
var best: Dictionary = scored[0]
|
||||
for i in range(1, scored.size()):
|
||||
var c: Dictionary = scored[i]
|
||||
var better: bool
|
||||
if absf(c.s - best.s) > 0.0001:
|
||||
better = c.s > best.s
|
||||
else:
|
||||
better = (c.id < best.id) if route != "max" else (c.id > best.id)
|
||||
if better:
|
||||
best = c
|
||||
return int(best.id)
|
||||
|
||||
|
||||
static func _on_step_row(row: Dictionary, _runner, st: Dictionary, collect_rows: bool) -> void:
|
||||
@@ -255,7 +340,8 @@ static func _on_node_entered(node: Dictionary, info: Dictionary, runner, st: Dic
|
||||
st.hp_before_rest = int(info.get("hp", 0)) - st.rest_healed
|
||||
runner.continue_after_node()
|
||||
elif t == "shop":
|
||||
runner.continue_after_node() # 商店房零购离开（GP+商店房批——自动路径直通 Boss；购买行为归冒烟商店专项断言/真人试玩）
|
||||
runner.continue_after_node() # 单关机制夹具（run_level）零购离开保持——smoke 商店/战斗断言面
|
||||
# 不掺购买行为；冒险口径（run_adventure）已在 _on_adventure_node_entered 改贪心买齐
|
||||
if logger != null:
|
||||
logger.log_node(node, info, runner.node_index)
|
||||
|
||||
@@ -274,9 +360,10 @@ static func _autoplay_battle(runner, settlement, st: Dictionary, logger) -> void
|
||||
if String(st.get("policy", "normal")) == "perfect":
|
||||
_autoplay_battle_perfect(runner, settlement, st, logger)
|
||||
return
|
||||
# 普通策略档 v2（2026-09-19 用户拍板）：基础定式画像——单步决策循环（move/dig/flag 经 settle_batch 提交，
|
||||
# 推理分级见 auto_player.gd 普通策略档节：新手级单数字满足规则 + 盲挖边界兜底、无 refresh）；
|
||||
# 旧"无推理挖最近格"口径随批退役。连续拒绝 ≥3 = 编排违约防御（同完美档 fail-fast 语义）。
|
||||
# 普通策略档 v3（2026-10-01 用户重定义）：完美策略 + 受控失误注入——v2「新手级单步推理+盲挖
|
||||
# 兜底」画像被真人实证推翻（零加成真人首通三关踩 23 雷 vs v2 普通档 L2 条件通关率 10.6%）。
|
||||
# v3 语义 = 会玩的普通玩家：推理与完美档同源（subset 解算/标旗/卡死刷新），失误只注入两手——
|
||||
# 标雷误挖（NORMAL_MINE_SLIP_P）+ 安全挖跳过（NORMAL_SKIP_DIG_P）。连续拒绝 ≥3 = 编排违约防御。
|
||||
var round_guard := 60
|
||||
while st.result == "" and runner.current_battle() == settlement and settlement.battle_result == "" and round_guard > 0:
|
||||
round_guard -= 1
|
||||
@@ -287,7 +374,11 @@ static func _autoplay_battle(runner, settlement, st: Dictionary, logger) -> void
|
||||
while runner.current_battle() == settlement and settlement.battle_result == "" and settlement.action_economy.phase == "Active" \
|
||||
and settlement.round_index == start_round and action_guard > 0 and rejects < 3:
|
||||
action_guard -= 1
|
||||
var act := AutoPlayer.novice_action(settlement.minesweeper)
|
||||
var act: Dictionary
|
||||
if bool(st.get("adv", false)):
|
||||
act = _apply_normal_slips(AutoPlayer.perfect_action(settlement.minesweeper, settlement), st) # v3：完美+失误
|
||||
else:
|
||||
act = AutoPlayer.novice_action(settlement.minesweeper) # 单关机制夹具：v2 新手画像保持
|
||||
var kind := String(act.get("kind", "stop"))
|
||||
if kind == "stop":
|
||||
break
|
||||
@@ -298,7 +389,13 @@ static func _autoplay_battle(runner, settlement, st: Dictionary, logger) -> void
|
||||
ok = _submit_intent(settlement, "dig", act.get("cell"), logger, start_round)
|
||||
elif kind == "flag":
|
||||
ok = _submit_intent(settlement, "flag_place", act.get("cell"), logger, start_round)
|
||||
elif kind == "refresh":
|
||||
ok = bool(settlement.player_request_refresh())
|
||||
if logger != null:
|
||||
logger.log_action("refresh", Vector2i(-1, -1), ok, {"round": start_round})
|
||||
if ok:
|
||||
if kind == "refresh":
|
||||
st.refreshes = int(st.refreshes) + 1
|
||||
rejects = 0
|
||||
else:
|
||||
rejects += 1
|
||||
@@ -308,9 +405,30 @@ static func _autoplay_battle(runner, settlement, st: Dictionary, logger) -> void
|
||||
logger.log_action("stop", Vector2i(-1, -1), _accepted(out_s), {"round": start_round})
|
||||
|
||||
|
||||
# 普通档 v3 失误注入（2026-10-01 用户规格）：只在完美策略的两类动作上出手——
|
||||
# ①标雷误挖：把「标旗已证雷」动作直接转成「挖该格」（真人手滑原型：知道是雷、位置点错，
|
||||
# 踩雷吃伤且该雷奖励报废）；期望踩雷增量 ≈ 每板已证雷数 × NORMAL_MINE_SLIP_P（目标完美+2/板）
|
||||
# ②安全挖跳过：把「挖已证安全格」动作转成本轮停手（少挖几个正确的雷——信息与推进双损失）
|
||||
# 失误流 = st.prng（策略流 RNG，种子派生自局种子——同种子逐位可回放；核心七流零消费零扰动）
|
||||
static func _apply_normal_slips(act: Dictionary, st: Dictionary) -> Dictionary:
|
||||
var prng: RandomNumberGenerator = st.get("prng")
|
||||
if prng == null:
|
||||
return act
|
||||
var kind := String(act.get("kind", "stop"))
|
||||
if kind == "flag" and prng.randf() < NORMAL_MINE_SLIP_P:
|
||||
st.mine_slips = int(st.get("mine_slips", 0)) + 1
|
||||
return {"kind": "dig", "cell": act.get("cell")}
|
||||
if kind == "dig" and prng.randf() < NORMAL_SKIP_DIG_P:
|
||||
st.skip_digs = int(st.get("skip_digs", 0)) + 1
|
||||
return {"kind": "stop"}
|
||||
return act
|
||||
|
||||
|
||||
# 完美策略档战斗驱动（2026-09-14）：单步决策循环——move/dig/flag 经 排雷玩法 settle_batch 提交，
|
||||
# refresh 经 settlement.player_request_refresh（战斗结算 核心编排：行动经济 扣 RAP + 退役 + 当场换新板），
|
||||
# stop 出内环走统一停手提交。连续拒绝 ≥3 = 编排违约防御（fail-fast 语义，停手防死循环）。
|
||||
# 2026-10-01 上界修正：选路改目标价值评分（_route_pick）+ 商店贪心买齐（_autoplay_shop）——
|
||||
# 「完美」须是元游戏意义上的上界（会购物会补给），不再只是扫雷子游戏上界。
|
||||
static func _autoplay_battle_perfect(runner, settlement, st: Dictionary, logger) -> void:
|
||||
var round_guard := 60
|
||||
while st.result == "" and runner.current_battle() == settlement and settlement.battle_result == "" and round_guard > 0:
|
||||
|
||||
Reference in new issue
Block a user