A Unified Model of Reasoning and Learning

Pei WangIWSSL 2021

Abstract

Traditional artificial intelligence has long maintained a strict dichotomy between reasoning—the application of formal logic to reach certain conclusions—and learning—the statistical process of improving performance through data. This paper introduces the Non-Axiomatic Reasoning System (NARS), a framework for Artificial General Intelligence (AGI) that posits reasoning and learning are not separate modules, but rather two facets of a single adaptive process. By operating under the Assumption of Insufficient Knowledge and Resources (AIKR), NARS provides a unified mathematical and logical foundation for systems that must function in real-time with incomplete information.

Introduction

The history of AI is characterized by a divide between logic-based "reasoning" and data-driven "learning." Reasoning systems, derived from mathematical logic, typically focus on deduction and theorem proving where truths are absolute and axioms are fixed. Conversely, machine learning focuses on function approximation and algorithmic optimization.

This separation presents a significant hurdle for the development of AGI. In real-world environments, an intelligent agent rarely possesses perfect axioms or infinite training data. Most contemporary attempts to bridge this gap involve "hybrid" systems that layer reasoning on top of learning. This paper proposes a more fundamental integration: a system where every inference step is simultaneously an update to the system’s internal model of the world, effectively treating reasoning as learning and learning as reasoning.

The NARS Framework

The core of the proposed model is the Non-Axiomatic Reasoning System (NARS). Unlike traditional systems that aim for absolute truth, NARS is built on the AIKR (Assumption of Insufficient Knowledge and Resources). This assumption dictates that the system must be:

  1. Finite: It has limited computational power and memory.
  2. Real-time: It must respond to tasks within deadlines.
  3. Open: It must be able to process novel information that contradicts its previous beliefs.

Experience-Grounded Semantics

In NARS, the "truth" of a statement is not a binary value (True/False). Instead, it is a measure of the degree to which the statement is supported by the system's past experience. Truth-values consist of two components:

  • Frequency: The proportion of positive evidence among all evidence seen.
  • Confidence: A value in the interval [0, 1) that indicates how much evidence the system has processed relative to a future-looking horizon.

Narsese and Term Logic

NARS utilizes a categorical language called Narsese. Unlike predicate calculus, which focuses on the properties of objects, Narsese is a term logic that focuses on the relationships between concepts (e.g., "S is a type of P"). This allows the system to perform flexible categorization and substitutability, which are essential for general intelligence.

Non-Deductive Inference

While classical logic prioritizes deduction, NARS treats induction (generalizing from specific instances), abduction (generating hypotheses to explain observations), and analogy as core components of its logical engine. These "weak" forms of inference allow the system to expand its knowledge base even when direct evidence is missing.

Unifying Reasoning and Learning

The most significant insight of the NARS model is the collapse of the distinction between reasoning and learning. In this framework:

  • Reasoning refers to the short-term, step-by-step derivation of a conclusion from premises.
  • Learning refers to the long-term, cumulative effect of these derivations on the system’s memory, priority distributions, and belief network.

Because NARS uses a priority-based system to manage its "concept network," it does not follow a fixed algorithm. Instead, it uses controlled concurrency to decide which tasks and beliefs to process based on their perceived importance and relevance. This means the system is always learning; every input is a task that can trigger forward inference (creating new beliefs) or backward inference (creating new goals).

Qualitative Results: The Logic in Practice

The paper demonstrates the system's capabilities through a qualitative walkthrough of conflicting information management, often illustrated by the "Tweety" example:

  1. Deduction: Given the premises "Robins are birds" and "Birds are animals," NARS derives "Robins are animals" with high frequency and confidence.
  2. Induction: Upon observing "Tweety is a robin" and "Tweety is yellow," the system tentatively induces that "Robins are yellow." This belief is held with lower confidence due to the limited sample size.
  3. Revision: If the system later receives evidence that "Tweety is NOT an animal," it does not experience a logical crash. Instead, it applies a revision rule to pool the conflicting evidence. It lowers its confidence in the previous chain of reasoning and updates its beliefs about the relationships between Tweety, robins, and animals accordingly.

This demonstrates a system capable of self-correction and "relative rationality"—making the best possible decision given its specific history and current resource constraints.

Limitations and Future Directions

Despite its theoretical elegance, NARS faces several challenges:

  • Implementation Complexity: The "Peewee Granularity" of NARS—breaking every process into tiny, flexible inference steps—requires highly sophisticated resource management that is difficult to scale to massive datasets.
  • Explainability at Scale: While individual inference steps are traceable and explainable, the emergent behavior of millions of concurrent priority-based inferences may become difficult for humans to interpret.
  • Scaling: Future work is focused on testing whether this unified strategy can outperform modular hybrid approaches in complex, real-world environments.

Conclusion

NARS represents a shift away from the pursuit of absolute, axiomatic truth in AI. By embracing the limitations of knowledge and resources, it provides a unified framework where learning is simply the long-term manifestation of reasoning. This approach offers a promising path toward more robust, adaptive, and human-like general intelligence.

Summary updated

BibTeX
@inproceedings{wang2021unified,
  author = {Pei Wang},
  title = {{A Unified Model of Reasoning and Learning}},
  booktitle = {IWSSL},
  year = {2021},
}
Figure