Analogy in AGISystem2 is geometric translation. If "king" is to "queen" as "man" is to "?", the answer comes from computing the offset between king and queen, then applying it to man. This transforms relational reasoning into vector arithmetic.

How It Works

The Reasoner handles analogy queries in three steps:

1. Compute offset – Calculate v(B) - v(A) to get the relational difference
2. Apply to source – Add offset to v(C) to get target vector
3. Find nearest – Use LSH to find the concept closest to the target

This works because similar relations produce similar offsets across different concept pairs. The "male→female" offset is approximately the same whether applied to king/queen, man/woman, or actor/actress.

Sys2DSL Syntax

# Basic analogy
ANALOGY king:queen :: man:?
# → woman

# With explicit relation
ANALOGY Paris:France :: Berlin:?
# → Germany (LOCATED_IN relation)

# Multi-step
ANALOGY doctor:hospital :: teacher:?
# → school (WORKS_AT relation)

Masked Analogy

Analogies can use masks to focus on specific dimensions:

Mask Type	Effect	Use Case
Full mask	All dimensions contribute	General analogies
Ontology only	Factual dimensions only	Scientific analogies
Custom domain	Specific dimension subset	Domain-specific transfer

Quality Metrics

Analogy results include confidence scores based on:

• Distance to target – How close is the answer to the computed point?
• Offset consistency – Does the same offset work for other known pairs?
• Relation preservation – Does the answer maintain the implicit relation type?

Related Documentation

• Reasoner – Analogy computation
• LSH – Finding nearest concepts
• Conceptual Spaces – Why vector arithmetic captures relations
• Deduction – Containment-based inference
• Induction – Learning concept boundaries