Difference between revisions of "Burnin"

Revision as of 20:03, 30 December 2005

Burnin (or Burn-in or Burning-in).

• Typically the chain will take some time before trees are sampled according to their probability
• Initially probability of trees increases with time
• Programmes need to be allowed to run until the probabilities are fluctuating randomly about a constant mean
• Data generated before the chain reaches a steadystate are discarded

• There are some disadvantages of picking trees near our current tree
• Our very first tree might not have a good P(D|T), and the first part of the chain might be in a valley. This will distort the calculation of posterior probabilities
• In practice: "burning-in" of the chain: start somewhere random, but run the process for a while, and discard the first part of the chain as atypical.