Analogy reasoning is a cognitive process that involves identifying similarities between two distinct domains or situations and using those similarities to infer further correspondences. It is a fundamental form of inductive reasoning, enabling knowledge transfer, problem-solving, and creative thinking. Formalized in fields such as logic, cognitive science, and artificial intelligence, analogy reasoning typically relies on mapping relational structures from a source domain to a target domain. Its applications range from legal argumentation and scientific discovery to machine learning and everyday reasoning.

1.1 Definition and scope

Analogy reasoning is the process of drawing a comparison between two different entities or systems—the source and the target—based on shared relational structures rather than mere surface attributes. It is a form of inductive inference that allows one to transfer knowledge from a familiar domain to an unfamiliar one. The scope of analogy reasoning extends across human cognition, formal logic, artificial intelligence, and many applied fields, making it a versatile tool for learning, explanation, and discovery.

1.2 Core components

1.2.1 Source domain and target domain

The source domain is the familiar or well-understood concept from which knowledge is drawn. The target domain is the less understood concept to which the analogy is applied. For example, in the classic analogy “atom is like a solar system,” the solar system is the source domain, and the atom is the target domain. The success of an analogy depends on the appropriateness of the mapping between these two domains.

1.2.2 Relational mapping

Relational mapping refers to the systematic alignment of relationships between elements of the source and target domains. It is the core of analogical inference, where correspondences are established not just between isolated objects but between higher-order relations (e.g., cause–effect, part–whole, or containment). For instance, in the atom–solar system analogy, the relation “planet orbits sun” maps to “electron orbits nucleus.”

1.2.3 Attributes versus relations

A crucial distinction in analogy theory is between attributes (surface features) and relations (structural connections). Surface attributes are specific properties of objects, such as color, size, or shape. Relations are abstract connections between objects, such as “larger than” or “revolves around.” Analogies are considered deeper when they rely on relational structure rather than surface similarity, though in practice both may contribute.

1.3 Types of analogies

1.3.1 Literal similarity

Literal similarity occurs when two domains share both surface attributes and relational structure to a high degree. For example, comparing one car to another of the same model involves literal similarity. These analogies are straightforward and usually not considered creative.

1.3.2 Surface similarity

Surface similarity refers to analogies based primarily on shared perceptual or superficial features, with little relational structure. For instance, comparing a ball to an orange because both are round and orange in color relies on surface similarity. Such analogies often mislead if used for reasoning.

1.3.3 Structural similarity

Structural similarity focuses on matching the underlying relational patterns between domains, even when surface features differ. The classic analogy “heart is like a pump” depends on structural similarity (pumping action) rather than appearance. Structural analogies are considered more powerful for knowledge transfer and are central to scientific and creative thinking.

2.1 Role in human reasoning

2.1.1 Deductive versus analogical reasoning

Deductive reasoning moves from general premises to specific conclusions with logical certainty, whereas analogical reasoning is inductive and probabilistic, drawing conclusions based on similarity between cases. Analogical reasoning is more flexible and context-sensitive, often used when no formal deductive rule is available. It is a common mode of everyday reasoning, enabling humans to learn from past experiences.

2.1.2 Developmental aspects (Piaget, Gentner)

Jean Piaget studied analogical reasoning as part of cognitive development, noting that young children often rely on surface similarity, while older children and adults become capable of mapping deeper relational structures. Dedre Gentner’s work on structure‑mapping theory further refined this view, showing that children as young as three can use relational analogies in simple contexts, but abstract relational reasoning develops gradually. The ability to inhibit surface features and focus on relations is a key milestone in cognitive maturation.

2.2 Cognitive models of analogy

2.2.1 Structure‑mapping theory (Gentner)

Proposed by Dedre Gentner in the early 1980s, structure‑mapping theory posits that analogical reasoning is a process of aligning the relational structures of source and target domains. The mapping is governed by systematicity—a preference for systems of interconnected relations rather than isolated features. The theory is implemented in the Structure‑Mapping Engine (SME) and has been influential in explaining how humans and machines can draw analogies.

2.2.2 Multi‑constraint theory (Holyoak & Thagard)

Developed by Keith Holyoak and Paul Thagard, the multi‑constraint theory argues that analogical mapping is guided by three interacting constraints: similarity (elements that look alike tend to map), structure (relational consistency), and purpose (the reason for drawing the analogy). These constraints are integrated in a parallel constraint satisfaction process, implemented in the ACME system. The theory accounts for pragmatic influences on analogical reasoning.

2.2.3 Copycat and analogy‑making (Hofstadter)

Douglas Hofstadter and his colleagues created the Copycat program, which models analogy‑making in micro‑domains (e.g., letter‑string analogies like “abc → abd” is like “pqr → ?”). The system represents analogies as dynamic, emergent structures built through bottom‑up and top‑down processes. Copycat emphasizes fluidity, creativity, and the role of perception in analogy formation, contrasting with more formal mapping theories.

3.1 Analogical proportion

3.1.1 Definition and properties

An analogical proportion is a relation of the form “A is to B as C is to D,” often written as A:B :: C:D. It expresses a structural equivalence between two ordered pairs. Formal definitions vary; one common approach treats analogical proportion as a quaternary relation satisfying symmetry (if A:B :: C:D then C:D :: A:B) and transitivity properties. Analogical proportions have been studied in logic, algebra, and linguistics.

3.1.2 Boolean analogies

Boolean analogies extend the concept to binary values (0 and 1) under logical operations. For example, an analogical proportion between four Boolean values can be defined using equality or difference constraints. These formalisms have been applied to reasoning tasks in artificial intelligence, such as analogy‑based learning and interpolation in logical frameworks.

3.2 Category‑theoretic models

Category theory provides a high‑level mathematical language for modeling analogies as functors or adjunctions between categories. In this view, a source domain and target domain are categories, and an analogy is a mapping that preserves certain structures (e.g., products, limits). Category‑theoretic analogies have been used to unify different domains in mathematics and science, and they offer a rigorous foundation for analogical inference.

3.3 Inductive logic programming approaches

Inductive logic programming (ILP) techniques can learn analogical rules from relational data. By representing domains as logic programs, ILP systems can find common relational patterns and transfer them to new contexts. For example, an ILP system might learn that “if an object rotates around another, the smaller one is attracted to the larger” and apply this to both celestial and atomic domains. These approaches combine symbolic reasoning with machine learning.

4.1 Overview of AI analogy systems

Computational analogy systems aim to automate the process of identifying and applying analogies. Historically, these systems have ranged from symbolic mapping engines to modern neural networks. The goal is to enable machines to transfer knowledge across domains, improve reasoning, and support problem‑solving. Despite advances, achieving human‑level flexibility in analogy remains an open challenge.

4.2 Early rule‑based systems

4.2.1 SME (Structure‑Mapping Engine)

The Structure‑Mapping Engine (SME), developed by Brian Falkenhainer, Ken Forbus, and Dedre Gentner, implements structure‑mapping theory. SME takes representations of source and target domains as semantic networks and outputs candidate mappings, focusing on matching relational structure with systematicity. It has been used in cognitive modeling and AI applications such as qualitative reasoning and educational software.

4.2.2 ACME (Analogical Constraint Mapping Engine)

ACME, created by Keith Holyoak and Paul Thagard, implements the multi‑constraint theory. It uses a parallel constraint satisfaction network to find the best mapping given similarity, structural, and pragmatic constraints. ACME has been applied to a variety of analogical reasoning tasks and has influenced subsequent connectionist models of analogy.

4.3 Modern machine learning approaches

4.3.1 Neural analogy solvers

Neural network approaches to analogy include models that learn to complete analogies in domains such as word‑pair (e.g., “king:queen :: man:?”) or visual patterns (e.g., Raven’s progressive matrices). These solvers often use specialized architectures like relational networks or memory‑augmented networks to capture pairwise relationships and induce the underlying mapping.

4.3.2 Embedding‑based analogical reasoning

Embedding methods map words, concepts, or entities into continuous vector spaces where analogical relationships can be represented as arithmetic operations. For example, in word embeddings (e.g., word2vec, GloVe), the vector offset “king – man + woman” approximates “queen.” This property is used to solve simple analogies, though it is limited to surface‑level relational patterns and may fail on complex structural analogies.

4.3.3 Large language models and analogies

Large language models (LLMs) such as GPT‑4 and LLaMA have demonstrated remarkable ability to produce and interpret analogies in natural language. They can generate analogical explanations, complete analogy puzzles, and even transfer knowledge across domains. However, their performance is based on statistical patterns from training data rather than explicit mapping mechanisms, leading to occasional errors and lack of robustness on novel or abstract analogies.

5.1 Scientific reasoning and discovery

Analogies have played a central role in scientific discovery. Famous examples include Johannes Kepler’s use of the analogy between light and sound to model optics, and the analogy between water flow and electric current that helped develop circuit theory. In modern science, analogies are used to generate hypotheses, design experiments, and communicate complex ideas.

Legal analogical reasoning involves comparing cases to determine whether similar legal rules apply. Common law systems rely on precedent, where a current case is evaluated by analogy to past cases. This process requires identifying relevant similarities and differences between fact patterns, assessing the strength of the analogy, and applying or distinguishing prior rulings.

5.3 Education and analogical learning

Educators use analogies to explain unfamiliar concepts by linking them to familiar ones. For example, the analogy “DNA is like a blueprint” helps students grasp the role of genetic information. Research shows that well‑designed analogies improve comprehension and retention, but they can also lead to misconceptions if the mapping is incorrect or overextended.

5.4 Problem‑solving and creativity

Analogical thinking is a cornerstone of creative problem‑solving. Designers, engineers, and artists often draw inspiration from one domain to solve problems in another—for instance, using bird flight as an analogy for aircraft wing design. Techniques such as “Think Different” or “Synectics” explicitly encourage analogical thought to generate novel solutions.

6.1 Ambiguity of analogy

Analogies are inherently open to interpretation. The same source‑target pair can yield multiple valid mappings, and different people may draw different conclusions. This ambiguity can be a weakness when precise reasoning is required, as in legal or scientific contexts. Determining which analogy is most appropriate often depends on background knowledge and goals.

6.2 Over‑extension and false analogies

Over‑extension occurs when an analogy is applied beyond its valid scope, leading to incorrect conclusions. False analogies are those that appear plausible but do not hold under scrutiny—for example, comparing the human body to a machine in a way that overlooks biological complexity. Recognizing and avoiding such pitfalls is essential for sound reasoning.

6.3 Computational complexity of mapping

Finding an optimal relational mapping between two domains is computationally challenging. Even for moderate‑sized representations, the number of possible correspondences can be huge, making exhaustive search infeasible. Heuristic algorithms (e.g., in SME and ACME) scale better but may miss optimal mappings. This complexity limits the performance of fully automated analogy systems, especially on large or noisy real‑world data.