anglican.bbvb — org.clojars.lsaldyt/anglican 1.0.1

*epsilon-one* clj

*epsilon-two* clj

add-gradient-to-proposal clj

(add-gradient-to-proposal q-dist gradient-map weights stepsize use-adagrad)

update proposals via gradient step

assoc-gradient clj

(assoc-gradient state address proposal value)

store in state the gradient value at an address

get-address clj

(get-address smp)

returns a unique identifier for sample checkpoint and the updated state

get-or-create-q! clj

(get-or-create-q! state address prior)

get the learned proposal at an address; initialize proposal with prior at new addresses.

get-variational clj

(get-variational state)

return the learned approximating distributions at each address

ignore? clj

(ignore? state dist)

determine whether to learn an approximation for a given distribution object

make-initial-state clj

(make-initial-state only exclude)

merge-q! clj

(merge-q! state proposals)

force specific approximating distributions: used for initialization, or to use as learned proposals within an importance sampler.

optimal-scaling clj

(optimal-scaling f g)

given two vectors f=wg and g, estimate Cov(f,g)/Var(g). this gives the optimal scaling for the variance reduction term.

proposal-gradient clj

(proposal-gradient gradient-samples log-weights)

compute the gradient (and appropriate weightings) at each address

update-proposals! clj

(update-proposals! particles stepsize use-adagrad)

outer loop of gradient update procedure to update proposal dist atom