Alexander Joossens

Machine Learning

In the theoretical course “Machine Learning” the following concepts were seen:

Part 1: Introduction: basic terminology and concepts; nearest neighbor approaches, evaluation

Part 2: Learning based on discrete search spaces

  • decision trees: basic algorithm, heuristics, pruning strategies, missing value handling, multi-target trees
  • rule learning and inductive logic programming: search in a refinement lattice, subsumption, least general generalization
  • ensemble methods: boosting, bagging, random forests, bias-variance tradeoff in ensembles
  • automata, learning theory: algorithms for learning automata, sample complexity, VC-dimension, Rademacher complexity

Part 3: Learning based on numerical optimization and search in continuous spaces

  • Concepts of statistical learning: learning as optimization, loss functions, regularization
  • Artificial neural networks: basic principles, specific network structures, convolutional neural nets, recurrent neural nets, auto-encoders, generative adversarial networks
  • support vector machines: basic principles, SMO algorithm, kernels, sequence / tree / graph kernels
  • dimensionality reduction
  • methods based on matrix factorization

Part 4: Probabilistic models

  • overview of PGMs
  • inference and learning in these models
  • statistical relational learning: Problog, Markov logic, …
  • Reinforcement learning: state-action spaces, problem setting, Q-learning, deep reinforcement learning

In the practical course “Machine Learning”, AlphaZero (a ML program to master the games of chess, shogi and go) was used to solve the game Dots and boxes.

AlphaGo employs neural networks and Monte Carlo Tree Search (MCTS). Neural networks predict moves and evaluate board positions, guiding MCTS exploration. The policy network suggests strong moves, while the value network estimates position strength. AlphaGo learns through supervised and reinforcement learning from human and self-play games. These strategies enable it to understand complex Go positions and make strategic decisions that surpass human expertise.

AlphaZero project

In the theoretical course “Data Mining”, many more ML techniques were seen.

Data Mining