accumulated_target

accumulated_target#

hybrid_aif.accumulated_target(target, durations)[source]#

A per-tick target as the conditional target of a step, one distribution per level.

durations is a positive (levels,) array. A per-tick Gaussian (loc, chol) with precision L becomes a ConditionalGaussian (parent_ndims=1) with the same loc and precision D_k L at level k (chol_k = chol / sqrt(D_k)). A per-tick Categorical becomes a ConditionalCategorical whose table has the child axis first and the level axis last, column k being log_softmax(D_k log_softmax(logits)).

Each level is a normalised density with its own normaliser, scored once: the cross-entropy against level k is D_k times the per-tick one up to a constant of that level, and that constant is not a cost the model sets. A fixed cost per option is the control node’s own target, never this helper’s. Design: docs/design/step-durations.md T4, T5.