- generic perturb
- make high order distributions real distributions
- allow to compose models: if X is a model we should be able in another model to write
x ~ X - smc
- dependency graph to allow optimized perturb
- should find a way to store and display sampling statistics for each sampler