Spaces:
Sleeping
Sleeping
File size: 11,439 Bytes
6f44ddb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 | import sys
from collections import defaultdict
def patch_file(filepath):
with open(filepath, 'r', encoding='utf-8') as f:
content = f.read()
# 1. Add attention_history to train_phase
old_train_init = """ action_history = []
consecutive_conf_increases = 0
consecutive_att_drops = 0
locked_phase_timer = 0
last_reward = 0.0"""
new_train_init = """ action_history = []
attention_history = []
consecutive_conf_increases = 0
consecutive_att_drops = 0
locked_phase_timer = 0
last_reward = 0.0"""
content = content.replace(old_train_init, new_train_init)
# 2. Add attention_history update in train_phase loop
old_train_loop = """ # Update tracking variables
action_history.append(action_idx)
if len(action_history) > 3:
action_history.pop(0)
if obs.confusion > prev_conf:
consecutive_conf_increases += 1
else:
consecutive_conf_increases = 0
if obs.attention < prev_att:
consecutive_att_drops += 1
else:
consecutive_att_drops = 0"""
new_train_loop = """ # Update tracking variables
action_history.append(action_idx)
if len(action_history) > 4:
action_history.pop(0)
attention_history.append(obs.attention)
if len(attention_history) > 4:
attention_history.pop(0)
if obs.confusion >= prev_conf:
consecutive_conf_increases += 1
else:
consecutive_conf_increases = 0
if obs.attention < prev_att:
consecutive_att_drops += 1
else:
consecutive_att_drops = 0"""
content = content.replace(old_train_loop, new_train_loop)
# 3. Add attention_history to evaluate
old_eval_init = """ action_history = []
consecutive_conf_increases = 0
consecutive_att_drops = 0
locked_phase_timer = 0
last_reward = 0.0"""
new_eval_init = """ action_history = []
attention_history = []
consecutive_conf_increases = 0
consecutive_att_drops = 0
locked_phase_timer = 0
last_reward = 0.0"""
content = content.replace(old_eval_init, new_eval_init)
# 4. Add attention_history update in evaluate loop
old_eval_loop = """ # Update tracking variables
action_history.append(action_idx)
if len(action_history) > 3:
action_history.pop(0)
if obs.confusion > prev_conf:
consecutive_conf_increases += 1
else:
consecutive_conf_increases = 0
if obs.attention < prev_att:
consecutive_att_drops += 1
else:
consecutive_att_drops = 0"""
new_eval_loop = """ # Update tracking variables
action_history.append(action_idx)
if len(action_history) > 4:
action_history.pop(0)
attention_history.append(obs.attention)
if len(attention_history) > 4:
attention_history.pop(0)
if obs.confusion >= prev_conf:
consecutive_conf_increases += 1
else:
consecutive_conf_increases = 0
if obs.attention < prev_att:
consecutive_att_drops += 1
else:
consecutive_att_drops = 0"""
content = content.replace(old_eval_loop, new_eval_loop)
# 5. Replace select_action function
# Because of how complex it is, we will locate its start and end
start_str = "def select_action("
end_str = "# 5. Dataset Loader"
start_idx = content.find(start_str)
end_idx = content.find(end_str)
new_select_action = '''def select_action(
q_table: defaultdict,
state: tuple,
epsilon: float,
rng: random.Random,
obs_attention: float,
obs_confusion: float,
consecutive_conf_increases: int,
consecutive_att_drops: int,
locked_phase_timer: int,
action_history: list[int],
attention_history: list[float],
last_reward: float
) -> int:
"""Strict hierarchical action selection for Guided RL following policy constraints."""
c, a, m, p, la, ps, ssi = state
misc_str = "none"
for k, v in MISCONCEPTION_MAP.items():
if v == m:
misc_str = k
break
allowed = list(ACTIONS.keys())
# Helper function to mask out actions
def mask_action(action_name: str):
idx = ACTION_TO_IDX.get(action_name, -1)
if idx in allowed:
allowed.remove(idx)
def mask_except(action_names: list[str]):
idxs = [ACTION_TO_IDX[name] for name in action_names]
allowed[:] = [act for act in allowed if act in idxs]
def force_action(action_name: str):
idx = ACTION_TO_IDX[action_name]
allowed[:] = [idx]
# --- Step 1: Calculate Action Diversity & History ---
action_counts = defaultdict(int)
for act in action_history:
action_counts[act] += 1
worked_example_idx = ACTION_TO_IDX["worked_example"]
we_used_consecutively = len(action_history) >= 2 and action_history[-1] == worked_example_idx and action_history[-2] == worked_example_idx
# --- Step 2: "Engagement Floor Rule" (from feedback) ---
att_below_3_count = sum(1 for att in attention_history if att < 3.0)
engagement_floor_broken = (consecutive_att_drops >= 3) or (att_below_3_count >= 2)
if engagement_floor_broken:
# Switch away from dominant action
if action_history:
dominant_action = max(action_counts, key=action_counts.get)
if dominant_action in allowed:
allowed.remove(dominant_action)
# To ensure switch, we force exploration if it was worked_example
if dominant_action == worked_example_idx:
mask_except(["explain", "question", "analogize"])
# --- Step 3: Hard Constraint 1 - Attention Safety ---
# If attention < 2.5: NO worked_example, prefer explain/question
if obs_attention < 2.5:
mask_action("worked_example")
if not engagement_floor_broken:
mask_except(["explain", "question"])
# If attention < 3.0: immediate strategy shift required
if obs_attention < 3.0:
if action_history and action_history[-1] in allowed:
allowed.remove(action_history[-1])
# If attention is falling for 2 consecutive steps -> switch strategy class entirely
if consecutive_att_drops >= 2:
if action_history and action_history[-1] in allowed:
allowed.remove(action_history[-1])
# --- Step 4: Hard Constraint 2 - Action Diversity ---
# No action type may exceed 60% frequency (in window of 4, max 2 occurrences)
for act, count in action_counts.items():
if count >= 2:
if act in allowed:
allowed.remove(act)
# No repeated worked_example more than 2 consecutive times.
if we_used_consecutively:
mask_action("worked_example")
# If allowed is empty because of diversity, force exploration action
if not allowed:
allowed = list(ACTIONS.keys())
mask_except(["question", "analogize"])
if not allowed: # If those were masked, fallback
allowed = [ACTION_TO_IDX["question"], ACTION_TO_IDX["analogize"]]
# --- Step 5: Hard Constraint 3 - Confusion Reduction Rule ---
# If confusion is NOT decreasing for 2 consecutive steps
if consecutive_conf_increases >= 2:
if misc_str == "conceptual":
mask_except(["analogize", "question"])
elif misc_str == "factual":
force_action("correct_fact")
elif misc_str == "procedural":
if not we_used_consecutively and ACTION_TO_IDX["worked_example"] in allowed:
force_action("worked_example")
else:
mask_except(["explain", "question"])
elif misc_str == "transfer":
mask_except(["explain", "analogize"])
# --- Step 6: Hard Constraint 4 - Factual Misconception Rule ---
if misc_str == "factual":
# Primary actions: correct_fact (must lead), explain (support)
# worked_example only if confusion < 6
if obs_confusion >= 6.0:
mask_action("worked_example")
if len(action_history) == 0:
force_action("correct_fact") # Must lead
# --- Step 7: Domain Selection Policies (Priors if not overridden) ---
if len(allowed) > 1 and not (consecutive_conf_increases >= 2 or engagement_floor_broken or obs_attention < 3.0):
if misc_str == "conceptual":
mask_except(["explain", "analogize", "question"])
elif misc_str == "procedural":
mask_except(["worked_example", "explain", "question"])
elif misc_str == "factual":
mask_except(["correct_fact", "explain", "worked_example"])
elif misc_str == "transfer":
mask_except(["explain", "analogize", "worked_example", "question"])
if not allowed:
allowed = list(ACTIONS.keys()) # Safe fallback
if rng.random() < epsilon:
return rng.choice(allowed)
# Argmax over ALLOWED actions only
q_vals = [q_table[state][act] if act in allowed else -1e9 for act in range(N_ACTIONS)]
return int(np.argmax(q_vals))
# ---------------------------------------------------------------------------
'''
content = content[:start_idx] + new_select_action + content[end_idx:]
# 6. We must also fix the function call signatures for select_action
old_call_train = """ action_idx = select_action(
q_table, state, current_epsilon, rng,
obs.attention, obs.confusion,
consecutive_conf_increases, consecutive_att_drops,
locked_phase_timer, action_history, last_reward
)"""
new_call_train = """ action_idx = select_action(
q_table, state, current_epsilon, rng,
obs.attention, obs.confusion,
consecutive_conf_increases, consecutive_att_drops,
locked_phase_timer, action_history, attention_history, last_reward
)"""
content = content.replace(old_call_train, new_call_train)
old_call_eval = """ action_idx = select_action(
q_table, state, 0.0, rng, # epsilon=0 for evaluation
obs.attention, obs.confusion,
consecutive_conf_increases, consecutive_att_drops,
locked_phase_timer, action_history, last_reward
)"""
new_call_eval = """ action_idx = select_action(
q_table, state, 0.0, rng, # epsilon=0 for evaluation
obs.attention, obs.confusion,
consecutive_conf_increases, consecutive_att_drops,
locked_phase_timer, action_history, attention_history, last_reward
)"""
content = content.replace(old_call_eval, new_call_eval)
with open(filepath, 'w', encoding='utf-8') as f:
f.write(content)
patch_file("scripts/qlearning_pipeline.py")
|