It reminds me of Bayesian model sampling, where you have a distribution over possible weights and 'draw' a model from the distribution for each evaluation... A problem is that there may be interesting co-dependencies amongst the weights which independent sampling will have a hard time getting right.