Spaces:
Paused
Paused
File size: 7,878 Bytes
a52f96d |
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 |
"""Verify that Teacher Agent is actually learning and improving."""
import numpy as np
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).parent))
from train_teacher import train_teacher
from teacher_agent import TeacherAgent
from interfaces import StudentState
def verify_teacher_improves():
"""Verify teacher agent's reward increases over time."""
print("=" * 70)
print("VERIFYING TEACHER AGENT LEARNING")
print("=" * 70)
# Train teacher
print("\nTraining teacher for 500 iterations...")
history, teacher, student = train_teacher(num_iterations=500, verbose=False)
# Analyze rewards over time
rewards = np.array(history['teacher_rewards'])
# Split into early and late phases
early_rewards = rewards[:100]
mid_rewards = rewards[100:300]
late_rewards = rewards[300:]
early_avg = np.mean(early_rewards)
mid_avg = np.mean(mid_rewards)
late_avg = np.mean(late_rewards)
print(f"\nReward Analysis:")
print(f" Early (iter 0-99): {early_avg:.3f}")
print(f" Mid (iter 100-299): {mid_avg:.3f}")
print(f" Late (iter 300-499): {late_avg:.3f}")
# Check if teacher is learning
improvement = late_avg - early_avg
print(f"\n Improvement: {improvement:+.3f}")
if improvement > 0.2:
print(" ✅ Teacher is learning! (late rewards > early rewards)")
elif improvement > 0:
print(" ⚠️ Teacher shows slight improvement")
else:
print(" ❌ Teacher is NOT learning (rewards decreasing or flat)")
# Check if teacher is exploiting good actions
stats = teacher.get_statistics()
# Find best actions (highest average reward)
avg_rewards_per_action = []
for idx in range(len(stats['action_counts'])):
if stats['action_counts'][idx] > 0:
avg_reward = stats['action_rewards'][idx] / stats['action_counts'][idx]
count = stats['action_counts'][idx]
avg_rewards_per_action.append((idx, avg_reward, count))
avg_rewards_per_action.sort(key=lambda x: x[1], reverse=True)
print(f"\nTop 5 Actions by Average Reward:")
for i, (idx, avg_reward, count) in enumerate(avg_rewards_per_action[:5]):
action = teacher._index_to_action(idx)
print(f" {i+1}. {action.topic}-{action.difficulty}-{'R' if action.is_review else 'N'}: "
f"avg_reward={avg_reward:.3f}, count={count}")
# Check if teacher preferentially selects high-reward actions in late phase
print(f"\nAction Selection Analysis (Late Phase):")
late_actions = history['actions'][300:]
late_rewards_for_actions = history['teacher_rewards'][300:]
# Group by action
action_reward_map = {}
for action, reward in zip(late_actions, late_rewards_for_actions):
key = (action.topic, action.difficulty, action.is_review)
if key not in action_reward_map:
action_reward_map[key] = []
action_reward_map[key].append(reward)
# Get top actions by frequency in late phase
action_counts_late = {}
for action in late_actions:
key = (action.topic, action.difficulty, action.is_review)
action_counts_late[key] = action_counts_late.get(key, 0) + 1
sorted_actions = sorted(action_counts_late.items(), key=lambda x: x[1], reverse=True)
print(f" Most frequently selected actions in late phase:")
for i, ((topic, diff, review), count) in enumerate(sorted_actions[:5]):
avg_reward = np.mean(action_reward_map.get((topic, diff, review), [0]))
print(f" {i+1}. {topic[:3]}-{diff[:2]}-{'R' if review else 'N'}: "
f"count={count}, avg_reward={avg_reward:.3f}")
# Verify teacher is using learned information
print(f"\n" + "=" * 70)
print("VERIFICATION RESULTS:")
print("=" * 70)
checks_passed = 0
total_checks = 4
# Check 1: Rewards improve over time
if improvement > 0.1:
print("✅ Check 1: Teacher rewards improve over time")
checks_passed += 1
else:
print("❌ Check 1: Teacher rewards do not improve significantly")
# Check 2: Teacher tries all actions (exploration)
unique_actions = len([c for c in stats['action_counts'] if c > 0])
if unique_actions >= 25:
print(f"✅ Check 2: Teacher explores actions ({unique_actions}/30)")
checks_passed += 1
else:
print(f"❌ Check 2: Teacher doesn't explore enough ({unique_actions}/30)")
# Check 3: Teacher has some preference (exploitation)
top_action_freq = sorted_actions[0][1] if sorted_actions else 0
if top_action_freq > 20:
print(f"✅ Check 3: Teacher shows preference (top action selected {top_action_freq} times)")
checks_passed += 1
else:
print(f"❌ Check 3: Teacher doesn't show strong preference")
# Check 4: Student improves (teacher's goal)
student_early = np.mean(history['student_accuracies'][:100])
student_late = np.mean(history['student_accuracies'][300:])
student_improvement = student_late - student_early
if student_improvement > 0.1:
print(f"✅ Check 4: Student improves significantly ({student_early:.3f} → {student_late:.3f})")
checks_passed += 1
else:
print(f"❌ Check 4: Student doesn't improve much")
print(f"\nTotal: {checks_passed}/{total_checks} checks passed")
if checks_passed >= 3:
print("\n✅ TEACHER AGENT IS LEARNING AND IMPROVING!")
else:
print("\n⚠️ Teacher agent may need tuning")
print("=" * 70)
return checks_passed >= 3
def verify_ucb_algorithm():
"""Verify UCB algorithm is working correctly."""
print("\n" + "=" * 70)
print("VERIFYING UCB ALGORITHM")
print("=" * 70)
teacher = TeacherAgent(exploration_bonus=2.0)
# Test: Give some actions high rewards
from interfaces import TeacherAction
good_action = TeacherAction(topic='history', difficulty='easy', is_review=False)
bad_action = TeacherAction(topic='science', difficulty='hard', is_review=False)
# Give good action high rewards multiple times
for _ in range(10):
teacher.update(good_action, 10.0)
# Give bad action low rewards
for _ in range(10):
teacher.update(bad_action, 0.5)
# Teacher should prefer good action
from mock_student import MockStudentAgent
student = MockStudentAgent()
selections = []
for _ in range(50):
student_state = student.get_state()
action = teacher.select_action(student_state)
selections.append(action)
good_selections = sum(1 for a in selections if a.topic == 'history' and a.difficulty == 'easy' and not a.is_review)
good_rate = good_selections / len(selections)
print(f"\nGood action selection rate: {good_rate:.2f}")
if good_rate > 0.3:
print("✅ UCB algorithm is working (prefers high-reward actions)")
else:
print("❌ UCB algorithm may not be working correctly")
# Verify UCB scores
ucb_scores = teacher._compute_ucb_scores()
good_idx = teacher._action_to_index(good_action)
bad_idx = teacher._action_to_index(bad_action)
print(f"\nUCB Scores:")
print(f" Good action (history-easy-N): {ucb_scores[good_idx]:.3f}")
print(f" Bad action (science-hard-N): {ucb_scores[bad_idx]:.3f}")
if ucb_scores[good_idx] > ucb_scores[bad_idx]:
print("✅ UCB correctly ranks good action higher")
else:
print("❌ UCB ranking may be incorrect")
print("=" * 70)
if __name__ == "__main__":
# Verify UCB algorithm
verify_ucb_algorithm()
# Verify teacher improves
print("\n")
success = verify_teacher_improves()
sys.exit(0 if success else 1)
|