Georgia Tech OMSCS Projects Overview

---

🔄 Machine Learning for Trading (ML4T)

• Martingale Strategy
  • Simulate American roulette to study statistical outcomes.
  • Focus: statistics, experimentation, plotting, reporting.
• Portfolio Optimization
  • Use scipy.optimize to create optimal stock portfolios.
  • Skills: financial data wrangling, constraints modeling.
• Assess Learners
  • Implement decision trees, random forests, and bagging.
  • Compare overfitting, underfitting, and learner behavior.
• Defeat Learners
  • Design datasets to “break” learners like linear regression vs trees.
• Market Simulator
  • Create a simulator that handles trades and computes portfolio performance.
• Indicator Evaluation
  • Evaluate technical indicators (SMA, BBP) and build a theoretical strategy.
• Q-Learning Robot
  • Use RL to navigate environments (used mainly for skill dev).
• Strategy Evaluation
  • Build rule-based and ML-based trading strategies and compare performance.

---

🤖 AI for Robotics (AI4R)

1. Hopscotch – Kalman Filters
   Navigate asteroids using Kalman filters for state prediction.

2. Solar System – Particle Filters
   Monte Carlo localization of a spacecraft using planetary gravity.

3. Drone Control – PID Controller
   Use PID to stabilize a drone for inspection tasks.

4. Warehouse – A* Search
   Automate warehouse logistics with efficient pathfinding.

5. Indiana Drones – SLAM
   Use SLAM to map dense jungle terrain with sensor data.

---

🤯 Machine Learning

• A1 – Supervised Learning
  SVM, KNN, RF on imbalanced datasets. Emphasis on feature analysis and model behavior.

• A2 – Randomized Optimization
  RHC, SA, GA, MIMIC on TSP, NN training, and other tasks. Focus on algorithm behavior and convergence.

• A3 – Unsupervised Learning
  PCA, ICA, ISOMAP, KMeans; visualizations and cluster interpretation.

• A4 – Reinforcement Learning
  Policy iteration, value iteration, Q-learning on Blackjack and Frozen Lake.

---

🧠 Deep Learning

• NN from Scratch
  Implemented ReLU/Softmax NN using only NumPy on MNIST. Focused on backprop and SGD.

• CNNs (From Scratch + PyTorch)
  Wrote conv/pool layers manually, then transitioned to PyTorch for deeper networks.

• Transformers & Sequence Models
  Implemented LSTM, Seq2Seq, Transformer for translation tasks.

• Edge Audio Classification (Group Project)
  Built classifier using CNNs and Transformers, distilled for edge deployment.

---