Machine learning for combinatorial optimization

Inputs $\to$ $ML(problem,parameters)$ $\to$ outputs
Collect all data from all generations to predict the next best move
Variants:

• Supervized learning: training data from disired result examples
• Reinforcement learning: training data from penalties/rewards exploration (random moves to explore, then exploit)

Embedding

Graph representation $\to$ numerical space vector (for each node/edge)

PageRank

Node (page) is important if many nodes (pages) point to it
Random walker walks along the graph, fitness on nodes based on number of visits

DeepWalk

Predicting neighboring words for randomly seelcted words - what is important for this word