Injecting Randomness

• different initial weights in ANN - didn't perform as well as bagging and cross-validated committees
• decision tree split criteria which chooses randomly among the best 20 tests at each node
• others used weighted random choice
• in ANN bootstrap sampling of training data and adding Gaussian noise to the input features
• Markov chain Monte Carlo method - transformation operators on hypothesis - *eventually* converges but usually after running a long while collect L classifiers and combined