Automatically select relevant training data from previous round

[kmzzhang]

For each round > 1, calculate the probability of previous round training parameters under the current round surrogate posterior. Include all previous training data within e.g. 5\sigma probability into next round training data.