Exponential Weights Algorithms

This is an important family of algorithms in Competitive On-line Prediction. At each trial the weight of each strategy in the benchmark class is multiplied by ⚠ $e^{-\eta l}$, where ⚠ $\eta$ is a constant called the learning rate and ⚠ $l$ is the strategy's loss. The master's prediction is obtained as a weighted average (in different senses) of the strategies' predictions.

Algorithms in this class include:

• Bayesian merging scheme
• Weighted Majority Algorithm
• Aggregating Algorithm
• Hedge Algorithm
• Weighted Average Algorithm
• Exponentiated Gradient