NEWS

---

arf 0.2.5.9000

• Add mirai/mori parallel backend as an alternative to foreach/doParallel (#62)
  • Shares large read-only inputs (training data, forest, learned parameters) across workers via mori, lowering memory use in adversarial_rf(), forde(), forge(), expct(), and lik()
  • Enable with active mirai daemons or options(arf.backend), see ?arf-options
  • Parallel calls now report the backend in use once per session, including when parallel = TRUE finds no backend and runs sequentially; silence with options(arf.verbose = FALSE)
• Reduce peak memory of expct() by processing conditions in bounded blocks (16x lower on one large internal case, at 12-27% more run time; tune via options(arf.block_rows), see ?arf-options)
• Reduce memory and dispatch overhead in forde() (fused per-tree parameter pass, per-tree coverage) and lik() (per-batch reduction)
• Fix forge() and lik() failing under non-forking foreach adapters (doParallel PSOCK clusters, doFuture multisession) with "object not found" errors: parallel worker bodies now take all inputs as explicit arguments (#62)
• Fix darf(), rarf() and earf() ignoring a pre-trained arf or params passed via ...: they retrained and recomputed anyway, then failed with "formal argument 'params' matched by multiple actual arguments"
• Fix a crash when arf runs with parallel = TRUE inside a forked worker (future::plan(multicore), mclapply(), fork-based doParallel) while mirai daemons exist in the parent process: the child inherited the daemon connection and used it, which aborts the process since mirai is not fork-safe. Forked children now ignore the parent's daemons and use the foreach path; forcing options(arf.backend = "mirai") there errors with an explanation
• Fix lik() overestimating likelihoods for mixed (continuous and categorical) queries: leaves where one variable block had zero density contributed the other block's density instead of zero, so values were too high and batch-dependent, and the arf path disagreed with the slower path (#72)
  • Fix lik() assigning a repeated factor pattern the likelihood of the preceding row rather than of its match
  • Fix lik() erroring on purely categorical queries with duplicate rows and no arf ("column name 'obs' is not found")

arf 0.2.5 (2026-09-21)

• Behavior change: New mtry argument for adversarial_rf() with default max(2, floor(sqrt(p))) instead of ranger's floor(sqrt(p)), which gave mtry = 1 for fewer than 4 features (#59)
• Export sample_from_leaves() for intra-leaf marginal sampling
• Avoid fractional recycling of factor column indices in sample_from_leaves() (#63)
• Fix nomatch = "force" fallback in forge() and expct() with evidence_row_mode = "separate" when evidence rows match no leaf (requires finite_bounds != "no" in forde()): errored with data.table input, and expct() silently filled impossible rows with misaligned values and returned NA-padded evidence columns for valid rows (#67)

arf 0.2.4 (2025-02-24)

• Let verbose=FALSE silence (some) warnings

arf 0.2.3

• Add impute() function for direct missing data imputation with ARF
• Add one-line functions darf(), earf(), rarf()

arf 0.2.2

• Faster and vectorized conditional sampling
• Use min.bucket argument from ranger to avoid pruning if possible
• Option to sample NAs in generated data if original data contains NAs
• Stepsize in forge() to reduce memory usage
• Option for local and global finite bounds

arf 0.2.0 (2024-01-24)

• Vectorized adversarial resampling
• Speed boost for compiling into a probabilistic circuit
• Conditional densities and sampling
• Bayesian solution for invariant continuous data within leaf nodes
• New function for computing (conditional) expectations
• Options for missing data

arf 0.1.3 (2023-02-06)

• Speed boost for the adversarial resampling step
• Early stopping option for adversarial training
• alpha parameter for regularizing multinomial distributions in forde
• Unified treatment of colnames with internal semantics (y, obs, tree, leaf)