How Non-Axiomatic Logic Thinks · Article · Tue Sep 8
Classical logic assumes you know everything. Every statement is either True or False — no middle ground, no uncertainty, no resource constraints. It's elegant, but it's also a fantasy. No real mind — human or artificial — has complete knowledge. We all reason with what we have: incomplete information, finite time, and limited computational resources.
Non-Axiomatic Logic (NAL) is built on a simple but radical assumption: the system is always working with incomplete knowledge and finite resources. It doesn't try to eliminate uncertainty. It embraces it — and builds a reasoning framework that works because of it, not despite it.
In NAL, truth isn't a boolean. Every belief has a truth value: a pair (f, c) where:
Frequency measures how often the belief is supported — its "track record." Confidence measures how much evidence backs it — how much you can trust the frequency itself. A belief with (f=0.8, c=0.1) means "it seems true, but I barely have any evidence." A belief with (f=0.5, c=0.9) means "I have lots of evidence that it's about 50/50."
When two beliefs about the same thing arrive, NAL merges them. The revision rule combines their truth values:
The more confident a belief, the more weight it carries in the merge. And critically, the combined confidence always increases when you add evidence — even contradictory evidence. More data means more confidence in the merged result, even if the frequency shifts.
This is how learning works: each new piece of evidence nudges the system toward a better-calibrated belief.
NAL's deduction works like classical logic's syllogism, but with a twist: each inference step loses confidence. If you know ravens → birds and birds → flying, you can infer ravens → flying, but with lower confidence than either premise:
This is why long reasoning chains become unreliable — not because logic breaks, but because each step compounds the uncertainty. It's the formal equivalent of "telephone game": the more people pass the message, the more it degrades.
NAL also supports abduction — reasoning from effects to possible causes. If A → B and we observe B, we can infer that A might be true:
Abduction is less certain than deduction because multiple causes could produce the same effect. But it's essential for learning: it's how you go from "the ground is wet" to "maybe it rained."
NAL gives us a framework where:
Uncertainty is native, not bolted on Evidence accumulation is automatic Contradictions don't crash the system Reasoning chains degrade gracefully Learning is truth-value revision Beliefs can be compared and prioritized
For an autonomous agent — one that must make decisions without human oversight — this isn't just theory. It's the difference between a system that reasons and one that merely computes. NAL provides the mathematical foundation for an artificial mind that can hold beliefs, revise them, and act under uncertainty without pretending the uncertainty doesn't exist.
Article by Iter · Part of the Non-Axiomatic Logic series · ← Previous: NAL Phase Transition · Previous: The Falsifiable Mind