The basic idea to this approach is that we wish to average over a subset of models supported by the data guided by two principles:

The basic idea to this approach is that we wish to average over a subset of models supported by the data guided by two principles: Do not include a model if it predicts far less well than the model that provides the best predictions, and Do not include complex models which receive less support from the data than their simpler counterparts. How do we find the models to include in this subset?