Skip to contents

Build Per-Timestep LTM Count Tables with Online Update

build_online_ltm_timestep_counts.Rd

For ltm+/both+ models: records the LTM state BEFORE observing x[t] at each timestep, then updates the LTM with x[t]. This matches IDyOM's online-update semantics, where prediction at position t uses only the training corpus plus x[1:t-1], not x[t:T].

Usage

build_online_ltm_timestep_counts(
  x,
  N,
  alphabet,
  init_ltm,
  ltm_update_exclusion = FALSE,
  ltm_start_token = TRUE
)

Arguments

x

Character vector of events (the prediction sequence).

N

Maximum n-gram order.

alphabet

Character vector of all symbols.

init_ltm

List of data.tables (format of counts_ltm): the pre-trained LTM state before any events from x are observed.

ltm_update_exclusion

Logical. Apply update exclusion during online update.

ltm_start_token

Logical. If FALSE, skip NA-lag contexts (IDyOM-compatible).

Value

List of length N+1, each a data.table with columns index, context_id, Event, Ce, C, t, t1 — same format as STM tables, with index in 1..T.