修改task1的策略形式化

.gitignore

@@ -1,4 +1,4 @@
-.venv/
+.venv*/
 
 *.egg-info
 dist/
@@ -9,8 +9,13 @@ __pycache__/
 *.pyd
 *.log
 *.bak
+.coverage
 .DS_Store
 .idea/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.ipynb_checkpoints/
 
 self_use/
 .vscode/
@@ -19,5 +24,33 @@ self_use/
 
 CLAUDE.md
 .claude/
+.trae/
+
+# 临时调试脚本与备份
+_debug_*.py
+_debug_*.txt
+_debug_output.txt
+_old_*.txt
+_test_*.py
+_test_*.txt
+_*.npy
+_eval_*.json
+
+# 临时评测结果（保留正式的 eval_results.json）
+docs/Mathematical_logic/report/eval_*_quick.json
+docs/Mathematical_logic/report/eval_safe_*.json
+docs/Mathematical_logic/report/eval_t*_pure.json
+docs/Mathematical_logic/report/eval_task5_fix.json
+docs/Mathematical_logic/report/eval_robust_*.json
+
+# LaTeX temporary files
+*.aux
+*.bbl
+*.blg
+*.fdb_latexmk
+*.fls
+*.out
+*.toc
+*.synctex.gz
 
 

_debug_sprite.py

@@ -0,0 +1,48 @@
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+
+PROJECT_ROOT = Path(__file__).resolve().parent
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+
+from nesylink.env import make_env
+from nesylink.core.constants import ACTION_LABELS, ACTION_LEFT, ACTION_RIGHT, ACTION_UP, ACTION_DOWN
+from student_agent.baseline_policy import PLAYER_COLORS
+
+
+def green_bounds(frame):
+    xs, ys = [], []
+    for color in PLAYER_COLORS:
+        mask = np.all(frame == color, axis=2)
+        yy, xx = np.nonzero(mask)
+        if len(xx) > 0:
+            xs.append(xx); ys.append(yy)
+    xs = np.concatenate(xs); ys = np.concatenate(ys)
+    return int(xs.min()), int(xs.max()), int(ys.min()), int(ys.max()), int(xs.mean()), int(ys.mean())
+
+
+def run():
+    env = make_env(task_id="mathematical_logic/task_1", observation_mode="pixels")
+    obs, info = env.reset(seed=0)
+    # move right a few ticks, then left, then up, then down - observe green bounds vs engpos
+    plan = [("right", 3), ("left", 6), ("up", 3), ("down", 6)]
+    action_map = {"right": ACTION_RIGHT, "left": ACTION_LEFT, "up": ACTION_UP, "down": ACTION_DOWN}
+    pos = info["agent"]["position_px"]
+    xmin, xmax, ymin, ymax, xmean, ymean = green_bounds(obs)
+    print(f"init engpos={pos} green xmin={xmin} xmax={xmax} ymin={ymin} ymax={ymax} mean=({xmean},{ymean}) left_min={xmin-4} left_mean={xmean-8}")
+    for direction, ticks in plan:
+        a = action_map[direction]
+        for _ in range(ticks):
+            obs, r, term, trunc, info = env.step(a)
+            pos = info["agent"]["position_px"]
+            xmin, xmax, ymin, ymax, xmean, ymean = green_bounds(obs)
+            print(f"{direction} engpos=({pos[0]:.0f},{pos[1]:.0f}) green x[{xmin},{xmax}] y[{ymin},{ymax}] mean=({xmean},{ymean}) left_min={xmin-4} left_mean={xmean-8}")
+    env.close()
+
+
+if __name__ == "__main__":
+    run()

_debug_task12.py

@@ -0,0 +1,66 @@
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+
+PROJECT_ROOT = Path(__file__).resolve().parent
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+
+from nesylink.env import make_env
+from nesylink.core.constants import ACTION_LABELS
+from student_agent.baseline_policy import Policy
+
+
+def run(task_id: str, seed: int = 0, max_steps: int = 1000):
+    env = make_env(task_id=task_id, observation_mode="pixels")
+    obs, info = env.reset(seed=seed)
+    policy = Policy()
+    policy.reset(seed=seed, task_id=task_id)
+    print(f"=== {task_id} seed={seed} ===")
+    prev_tile = None
+    for step in range(1, max_steps + 1):
+        action = policy.act(obs, info)
+        tile = policy.history.facing if False else None
+        # detect player tile from obs
+        pt = policy.detect_player_tile(obs)
+        keys = int(info.get("inventory", {}).get("keys", 0))
+        label = ACTION_LABELS[action] if action < len(ACTION_LABELS) else str(action)
+        monster = policy.detect_monster_tile(obs) if task_id.endswith("task_2") else None
+        agent = info.get("agent", {})
+        pos = agent.get("position_px", None)
+        det_px = policy.detect_player_px(obs)
+        last_msg = info.get("engine", {}).get("last_message") if "engine" in info else None
+        events = [r.get("name") for r in info.get("events", {}).get("records", [])]
+        prev_action = getattr(run, "_pa", None)
+        prev_pos = getattr(run, "_pp", None)
+        meaningful = (
+            pt != prev_tile
+            or action != prev_action
+            or events
+            or "blocked" in str(events)
+            or pos != prev_pos and (pos is None or prev_pos is None or abs(pos[0] - prev_pos[0]) + abs(pos[1] - prev_pos[1]) > 15)
+        )
+        marker = ""
+        if pt == prev_tile and action in (1, 2, 3, 4) and "blocked" in str(events):
+            marker = " <-- BLOCKED"
+        prev_tile = pt
+        run._pa = action
+        run._pp = pos
+        if meaningful or marker:
+            print(
+                f"step {step:3d} tile={pt} engpos={pos} detpx={det_px} keys={keys} facing={policy.history.facing} "
+                f"action={label} events={events} monster={monster}{marker}"
+            )
+        obs, reward, terminated, truncated, info = env.step(action)
+        if terminated or truncated:
+            print(f"  -> terminated={terminated} truncated={truncated} reason={info.get('terminal_reason')}")
+            break
+    env.close()
+
+
+if __name__ == "__main__":
+    task = sys.argv[1] if len(sys.argv) > 1 else "mathematical_logic/task_1"
+    run(task)