Intro to AI

A collection of projects from my Introduction to Artificial Intelligence course, covering multi-armed bandits, Markov decision processes, and neural network classifiers.

Projects

BanditProject

Implementation of the multi-armed bandit problem using Thompson Sampling with Dirichlet priors. Features social learning from other agents' choices and Monte Carlo simulation methods.

Key files:

• bandit.py - Thompson Sampling with Dirichlet priors
• NewBandit.py - Information-Directed Sampling (IDS) for optimal exploration-exploitation
• BayesUCB.py - Bayesian Upper Confidence Bound implementation
• monte_carlo.py - Monte Carlo simulation methods
• simulator.py - Bandit simulation environment

BucketsProject

A Markov Decision Process (MDP) solution to the classic water bucket problem. Given three buckets of different capacities, the goal is to find the optimal policy for reaching a target water level through fill, empty, and pour actions.

Key files:

• buckets_mdp.py - MDP implementation with value iteration

BlackjackProject

An MDP-based Blackjack solver with both infinite-deck analysis and finite-deck card counting. Uses policy iteration to compute optimal hit/stand decisions and includes an interactive tournament agent for live play assistance.

Features:

• Policy iteration for computing optimal strategy
• Adaptive MDP that recomputes optimal play based on remaining cards
• Card counting with exact deck tracking
• Monte Carlo simulation for strategy comparison (basic vs adaptive vs Hi-Lo counting)
• Tournament configuration with rank removal (e.g., no 5s)

Key files:

• blackjack_mdp.py - Infinite-deck MDP with policy iteration
• blackjack_game.py - Core game logic and hand evaluation
• adaptive_mdp.py - Finite-deck MDP that adapts to deck composition
• finite_deck_tracker.py - Exact card counting and probability tracking
• tournament_agent.py - Interactive CLI for live tournament play
• monte_carlo_simulator.py - Strategy comparison via simulation

Usage:

# Compute optimal strategy for infinite deck
python blackjack_mdp.py

# Run interactive tournament agent
python tournament_agent.py

# Compare strategies via Monte Carlo simulation
python monte_carlo_simulator.py

PerceptronProject

Image classification using three progressively complex neural network architectures:

1. Single Perceptron - Basic linear classifier
2. Multi-Layer Perceptron (MLP) - Feedforward neural network
3. Convolutional Neural Network (CNN) - Deep learning approach

Supports GPU acceleration on M1 Macs via Metal.

Key files:

• classifier1_perceptron.py - Single perceptron classifier
• classifier2_mlp.py - MLP classifier
• classifier3_cnn.py - CNN classifier

Usage:

# Train a model
python classifier2_mlp.py --train --epochs 50 --gpu

# Classify images
python classifier2_mlp.py -d [image_directory]