PLN Full

Historical Design Decisions (mailing-list-backed)

Rules-as-Atoms Architecture. The 2014 transition from C++ hardcoded PLN rules to AtomSpace-based BindLink rules was a deliberate architectural choice enabling three capabilities: (1) the Pattern Matcher can execute rules directly, (2) PLN can reason about its own rules (meta-inference), and (3) MOSES can evolve new inference rules. However, this immediately created a combinatorial explosion problem — running a Deduction BindLink on a large AtomSpace creates a "gazillion" new InheritanceLinks. The solution was ECAN's attentional focus: only atoms with sufficient Short-Term Importance participate in inference. Tournament selection (choosing \(k\) random atoms and taking the highest-STI one) was adopted as the specific selection mechanism. This architecture — rules-as-atoms with attention-gated combinatorics — defined the Unified Rule Engine design. (Rule engine discussion, Simplifying-the-PLN-rule-set-some-possibilities, 2014)

PLN vs. MLN (Markov Logic Networks). A 2014 bridge to the Tuffy MLN engine established the weight mapping: \(w = c \cdot \log\bigl(s / (1-s)\bigr)\), where \(s\) is PLN strength and \(c\) is confidence. Key finding: Tuffy was much faster for closed-domain batch inference, but PLN's lazy grounding — creating atoms only as needed rather than pre-materializing all possible groundings — avoids the combinatorial explosion that breaks MLN when the constant domain is large. This led to a hybrid strategy: MLN for closed-domain batch processing, PLN for open-world reasoning. The lazy grounding advantage is a fundamental architectural property of the AtomSpace that carries forward to MORK. (PLN-and-MLN, Cosmo Harrigan / Matt Iklé / Ben Goertzel, 2014)

Inference Control as Self-Referential Learning. Nil Geisweiller (2017) proposed and prototyped inference control learning: mining the history of past inference traces to discover patterns that speed up future inferences. The approach is explicitly self-referential — "ultimately you need an AGI to address that, but wait, we have one" — with PLN reasoning about its own inference histories to learn which rules to apply when. This is a direct precursor to the learned control policies in the current MORK-based design. (Inference-control-problem, Inference-meta-learning, 2017)

Backward Chaining as Forward Control. Nil Geisweiller (2016) observed that backward chaining can be embedded within the forward chainer's control mechanism: the backward propagation logic determines which forward steps to take next. This architectural insight — that backward chaining is reducible to goal-directed forward chaining control — anticipated the polyward chaining approach now used in the MeTTa implementation. (The-no-BC, 2016)

The "Reasoning Forests" Gap. A structural limitation acknowledged in the OpenCog-era PLN design: "reasoning can never be particularly deep" due to combinatorial explosion, and the system failed to provide "infrastructure for managing 'reasoning forests', or multiple, independent 'reasoning agents'" — a contrast with machine learning's decision-tree forests that run many independent estimators in parallel. Attention allocation was required to identify "fruitful directions of inference," but even with ECAN guidance, the single-threaded chaining architecture could not explore multiple reasoning paths simultaneously. This limitation directly motivated both the MORK parallel execution model and the geodesic inference control in the current design. (Provenance: official-site, wiki.opencog.org— PLN design notes)

System Interfaces and Dependencies

• ECAN: attention allocation guides which inferences are worth pursuing, preventing combinatorial explosion. PLN results inform which atoms deserve increased attention. (See Harrigan et al. 2014.)
• MOSES: PLN evaluates logical consistency of evolved programs; MOSES evolves inference control strategies for PLN.
• Pattern Mining: mined patterns become inference templates and proof search heuristics.
• Weakness Theory: the PLN-Quantale framework enables weakness-guided scheduling across both execution modes.
• Geodesic Control: maintains forward (\(f\)) and backward (\(g\)) factors simultaneously, choosing steps that maximize \(\Delta\log(f \times g)\) per unit cost.
• MetaMo: motivational framework determines which inference goals are most urgent.

Implementation Anchors

• PLN (primary) — consolidated lib_pln.metta on PeTTa. Rules: Revision, Modus Ponens, Deduction, Induction, Abduction, Evaluation, Negation. Configurable parameters. Imported via PeTTa git-import.
• pln-experimental — research platform with complete mathematical formalization (nuPLN.pdf), multiple rule representations (match-based, entailment-based, equality-based, dependent-types), ETV support, program synthesis experiments.
• chaining — DTL-based forward/backward/polyward chaining, Metamath proofs, modal logic, program verification.
• PLN2 (patham9/PLN) — fork preserving older modular src/ structure.
• PLN3 (ngeiswei/PLN) — fork with extended rule set (Symmetric Modus Ponens, Transitive Similarity).
• hyperon-pln — experimental Hyperon port exploring four implementation approaches with Idris2 prototyping.
• Legacy: opencog/pln(C++/Scheme, superseded)

Implementation Findings (discussion-backed, code-corroborated)

• Premise ordering performance: Moving the most specific premise first in modus ponens reduces proof search from minutes (Hyperon Experimental) to ~100ms (PeTTa) — a ~1000× speedup, demonstrating that inference control via premise ordering is a critical optimization even before formal attention allocation. (nil, PLN Development Mattermost, Oct 2025; PeTTa PR #21)
• STV/Beta distribution conversion: Code at pln-experimental/metta/common/truthvalue/TruthValue.metta provides conversion between PLN Simple Truth Values and Beta distribution parameters. STV strength maps to Beta mode (not mean); confidence relates to sample count via \(w = N/(N+k)\) where \(k\) is lookahead. The product of two Beta distributions is not Beta, but Beta approximation with matched mean/variance works well in practice. (nil, matt_ikle, robert.wuensche, PLN Development, Jun–Jul 2025)
• Noisy-OR truth function: PLN now supports Truth_NoisyOr alongside Truth_Revision for aggregating independent alternative causes (OR semantics vs. weighted-average revision). Achieves near-identical results to ProbLog on standard benchmarks including noisy-OR disease models. The default P(B|¬A) = 0.2 in modus ponens is now a settable parameter. (abdu, patham9, PLN Mattermost, Mar 2026)
• Virtual link performance cliff: In OpenCog Classic, wrapping queries in GetLink with virtual links (AbsentLink) caused a 1250× slowdown (4ms → 5s) because the pattern matcher attempted inference on ungrounded variables. The backward chainer never fully supported targets containing virtual links — a fundamental limitation that motivated the MeTTa redesign where pattern matching and rule chaining share a unified execution model. (Alexander Gabriel, Nil Geisweiller, opencog-ml, Mar 2020)
• PLN-based next token prediction: Experimental use case exploring PLN inductive/abductive reasoning for token prediction — potentially able to emerge ontological hierarchies from text sequences. Use case document at pln-experimental/docs/use-cases/next-token-prediction/. (nil, PLN Development, Jun 2025)
• Classic PLN performance baseline: The first working multi-step PLN forward inference (March 2016) took 4 days to run 1M inference steps and was non-deterministic due to atoms being sorted by memory address. This was the state of the art in OpenCog Classic — the context against which the ~1000× PeTTa speedup should be measured. (mailing-list-backed: A-pretty-sophisticated-working-PLN-forward-inference-thanks-to-Nil, 2016)