Skip to content

Projects 4: Learning

CS50's Introduction to Artificial Intelligence with Python

Project 1 — Shopping

Build a k-nearest-neighbor classifier to predict whether online shoppers will complete a purchase.

Background

When users browse a retail site, most won't buy. A naive classifier that always predicts "no purchase" would have high accuracy but be useless. This project emphasizes two metrics:

  • Sensitivity (true positive rate) — proportion of actual purchasers correctly identified
  • Specificity (true negative rate) — proportion of non-purchasers correctly identified

Data

~12,000 user sessions with 17 evidence columns + 1 label (Revenue: TRUE/FALSE):

Columns Type
Administrative, Informational, ProductRelated (counts) Integer
Duration columns, BounceRates, ExitRates, PageValues, SpecialDay Float
Month Integer (0=Jan … 11=Dec)
OperatingSystems, Browser, Region, TrafficType Integer
VisitorType 1 (returning) or 0 (other)
Weekend 1 (True) or 0 (False)

Specification

Implement 3 functions in shopping.py:

load_data(filename)

Returns (evidence, labels) tuple: - evidence — list of 17-element lists (numeric) - labels — list of integers (1 = purchase, 0 = no purchase) - Apply all type conversions from the table above

train_model(evidence, labels)

Returns a fitted KNeighborsClassifier with k=1.

from sklearn.neighbors import KNeighborsClassifier
model = KNeighborsClassifier(n_neighbors=1)
model.fit(evidence, labels)
return model

evaluate(labels, predictions)

Returns (sensitivity, specificity) as floats 0–1: 

sensitivity = correctly predicted purchases / total actual purchases
specificity = correctly predicted non-purchases / total actual non-purchases

Constraints

  • Only modify the three specified functions
  • Use Python's standard csv module for file loading
  • Don't modify shopping.csv
pip3 install scikit-learn
check50 ai50/projects/2024/x/shopping

Project 2 — Nim

Teach an AI to play Nim optimally through reinforcement learning (Q-learning).

Background

Nim: players take turns removing objects from piles. The player forced to take the last object loses.

  • State: current pile configuration, e.g., [1, 1, 3, 5]
  • Action: (i, j) = remove j objects from pile i
  • Rewards: +1 for winning, -1 for losing, 0 for continuing
  • Q-learning update:
Q(s, a) ← Q(s, a) + α × (reward + future_reward − Q(s, a))

Code Structure

Nim — game state, available_actions(piles), move execution (pre-built)

NimAI — learning agent with self.q (Q-value dict), alpha (learning rate), epsilon (exploration rate)

Training occurs through self-play simulations.

Specification

Implement 4 functions in NimAI:

get_q_value(state, action)

Returns Q-value for the state-action pair. Returns 0 if not in self.q. Convert state lists to tuples for dict keys.

update_q_value(state, action, old_q, reward, future_rewards)

Applies the Q-learning update: 

new_Q = old_Q + alpha × (reward + future_rewards − old_Q)
Stores result in self.q.

best_future_reward(state)

Returns the highest Q-value among available actions. Returns 0 if no actions available or all Q-values are unknown.

choose_action(state, epsilon=False)

  • Greedy (epsilon=False): Pick action with highest Q-value
  • ε-greedy (epsilon=True): With probability self.epsilon pick randomly; otherwise pick best action
  • Ties may resolve to any option

Hints

  • tuple(list) converts states for use as dict keys
  • After training (self-play), the AI should never lose to a human opponent
python play.py  # interactive testing after training
check50 ai50/projects/2024/x/nim