Q&A — Phase 26.4 CommonSenseEngine: Knowledge Sources & Inference Limits

[web3guru888]

Frequently Asked Questions — CommonSenseEngine

Q1: Where does the initial common-sense knowledge come from?

A: Three sources:

1. Bootstrap assertions: Pre-loaded basic facts (physical properties, typical behaviors, causal chains). Curated, high-confidence.
2. Experience-based acquisition: As the agent interacts with the world, observed regularities become assertions. Lower initial confidence, strengthened by repetition.
3. Analogical transfer: When encountering a new concept, the engine infers default properties from similar known concepts via ConceptGraph similarity.

Q2: How does multi-hop inference maintain confidence?

A: Confidence decays multiplicatively per hop with a configurable decay factor:

conf_chain = ∏(conf_i) × decay^num_hops

With decay=0.85 and 3 hops of confidence 0.8 each: 0.8³ × 0.85³ = 0.512 × 0.614 = 0.315. This prevents long, unreliable chains from appearing confident.

Q3: How is plausibility different from probability?

A: Plausibility scores are qualitative assessments, not calibrated probabilities. A score of 0.7 means "quite plausible based on available evidence" — not a 70% chance of being true. The score is computed from evidence signals, not from frequency data. This follows Collins & Michalski's (1989) framework where plausibility depends on certainty of premises, typicality of instances, and similarity of domains.

Q4: What are ExpectationFrames used for?

A: ExpectationFrames serve two purposes:

1. Prediction: Given a context (e.g., "restaurant dinner"), generate expected events (enter → sit → order → eat → pay → leave). Used by SocialPredictor (24.4) for behavioral anticipation.
2. Surprise detection: When an observed event violates the expected script, it triggers SurpriseDetector (13.4). Violations above violation_threshold (0.3) generate surprise signals for attention allocation.

Q5: How does analogical transfer differ from AnalogicalReasoner (20.2)?

A: AnalogicalReasoner (20.2) performs structural analogy mapping (Gentner's structure mapping theory) — finding deep relational correspondences between domains. CommonSenseEngine's analogical transfer is simpler — it transfers properties from similar concepts based on ConceptGraph similarity scores. Think of it as shallow analogy (property transfer) vs. deep analogy (relational mapping).

Q6: Can common-sense assertions be retracted?

A: Yes. Assertions have a confidence field that decays over time if not reinforced. Assertions contradicted by direct evidence have their confidence reduced via the plausibility scoring mechanism. Assertions that drop below a minimum confidence threshold (0.05) are garbage-collected during maintenance cycles.

Q7: How does CommonSenseEngine handle cultural/contextual variation?

A: The context field on CommonSenseAssertion allows situational qualifiers. For example, "eating with hands" might have context="Indian cuisine" vs. "eating with fork" with context="Western dining". When querying, the engine considers context matching — assertions with matching or null context are preferred. This is inspired by Cyc's microtheory approach (Lenat 1995).

---

Issue: #581 | Planning: #577