Abduction in AGISystem2 is reverse geometric search. Given an observation (effect), the system finds concepts that could explain it by searching backwards through relations. Instead of "X causes what?", abduction answers "what causes X?".

How It Works

The Reasoner handles abductive queries by:

1. Apply inverse permutation – For "? CAUSES Smoke", use π⁻¹(CAUSES) to reverse the relation encoding
2. Build search vector – Create a query vector from the observation using the inverted permutation
3. LSH retrieval – Use LSH to find candidate concepts near the query vector
4. Rank by distance – Score candidates by masked L1 distance

Example Queries

# What causes smoke?
ASK "? CAUSES Smoke"
# → Fire (0.92), Volcano (0.45), Engine (0.38)

# What is located in France?
ASK "? LOCATED_IN France"
# → Paris (0.95), Lyon (0.88), Marseille (0.85)

# What treats headache?
ASK "? TREATS Headache"
# → Aspirin (0.91), Ibuprofen (0.87), Rest (0.65)

Comparison with Deduction

Aspect	Deduction	Abduction
Direction	Premise → Conclusion	Observation → Explanation
Question	"Does X follow from Y?"	"What explains Y?"
Permutation	Forward π	Inverse π⁻¹
Certainty	Guaranteed	Probabilistic (multiple candidates)

Use Cases

• Diagnosis – Given symptoms, find possible diseases
• Debugging – Given an error, find possible causes
• Explanation – Given an observation, find reasons
• Root cause analysis – Trace backwards through causal chains

Related Documentation

• Reasoner – Query processing for abduction
• Permutation Binding – How inverse permutations work
• LSH – Candidate retrieval for abductive search
• Deduction – Forward inference
• Induction – Learning from examples