Introduction
Most mainstream artificial intelligence systems are built on the assumption that they will have sufficient data, time, and computational power to solve specific classes of problems. However, real-world intelligence—whether human or artificial—must often act when knowledge is incomplete, information is contradictory, and deadlines are tight. This paper provides a high-level overview of the Non-Axiomatic Reasoning System (NARS), a unified model of intelligence designed to operate under the "Assumption of Insufficient Knowledge and Resources" (AIKR). Rather than following fixed algorithms for narrow tasks, NARS treats intelligence as the ability to adapt to an unpredictable environment using experience-grounded reasoning and self-organizing memory.
The Core Constraint: AIKR
The fundamental challenge for artificial general intelligence (AGI) is the AIKR constraint. Traditional "logicist AI" often failed not because of logic itself, but because of the use of axiomatic logic (such as predicate calculus), which assumes absolute truths and infinite resources. NARS proposes a "non-axiomatic" alternative. In this framework, intelligence is not a collection of specialized functions—like separate vision, planning, or language modules—but a single, unified principle of adaptation to a complex world with limited means.
Non-Axiomatic Logic and Experience-Grounded Semantics
NARS differs fundamentally from both symbolic AI and connectionist approaches through its unique logical and semantic structure:
- Non-Axiomatic Logic (NAL): NARS utilizes a categorical logic similar to Aristotle’s term logic. This allows the system to perform not just deduction, but also induction, abduction, and analogy. Reasoning is viewed as a process of substituting and constructing concepts to summarize experience.
- Experience-Grounded Semantics: Truth-values in NARS are not binary (True/False). Instead, they are numerical measurements of how much a statement agrees with the system's past experience. These values consist of frequency (the ratio of positive vs. negative evidence) and confidence (the estimated effect of future evidence on the current frequency).
- Concept-Centered Representation: Instead of symbols representing external objects, NARS uses "terms" organized into concepts. The meaning of a concept is defined entirely by its experienced relations to other concepts, such as inheritance, similarity, or implication.
System Architecture and Problem Solving
NARS does not use fixed algorithms for problem-solving. Instead, it functions as an "anytime algorithm," distributing resources among tasks based on priority. It reports the best solution found within the given time constraints and continues to look for better ones if resources allow. This self-organizing memory ensures the system remains open to unanticipated tasks and can manage finite storage in real time.
Key Insights: Intelligence as Adaptation
In the NARS framework, several traditional AI problems are reframed:
- Learning as a Side Effect: Learning is not a separate module or training phase. It is the long-term effect of the reasoning process itself. As the system reasons to solve problems, it builds new conceptual relations and adjusts priorities, effectively learning from its own activity.
- Solving the Problem of Induction: By grounding validity in past experience rather than future truth, NARS bypasses the philosophical hurdles of induction. An inference is considered "valid" if it correctly summarizes what the system has seen so far, even if future evidence eventually contradicts it.
- Unified Cognition: Functions like planning, perceiving, and predicting are not separate algorithms but different perspectives of the same underlying reasoning process.
Conclusion and Future Directions
NARS provides a coherent, unified theory of AGI that avoids the "symbol grounding problem" by defining meaning through internal relations. It provides a formal way to handle uncertainty and operates effectively under strict real-time resource constraints.
While the current framework distinguishes between learned object-level knowledge and built-in meta-level mechanisms (like inference rules), future work explores how these meta-rules might evolve across generations of systems. Additionally, the local nature of NARS inference suggests significant potential for hardware acceleration and distributed processing, moving closer to a truly adaptive, general-purpose intelligence.
Summary updated