TL;DR
This paper provides a theoretical framework for understanding why similarity is a good heuristic for analogical reasoning, moving beyond the traditional assumption that the most similar analogue is the most likely to yield correct solutions. By statistically analyzing the probability of a successful analogy based on attribute similarity, the authors demonstrate a quantitative relationship that aligns with empirical data from human and animal studies.
What Problem Does This Paper Address?
The study of analogy has long relied on the intuition that if two things are similar in some ways, they are likely to be similar in others. This "analogy by similarity" approach suggests that the best analogue for a given problem (the target) is the one that shares the most characteristics with it (the source). However, this intuition has lacked a rigorous justification, and the relationship between the degree of similarity and the probability of a successful analogy has remained unclear. Furthermore, previous work, like Tversky's, has shown that similarity itself is a complex, context-dependent measure, making its direct use in analogy problematic. This paper aims to bridge this gap by providing a quantitative analysis that links similarity measures to the probability of analogical success, offering a theoretical foundation for this common heuristic.
Methods
The authors propose a simplified model of analogy based on attribute-value representations. They define similarity as the number of matching attribute values between a source and a target. The core of their method involves calculating the probability of a successful analogy, denoted as $p(d, r)$, where $d$ is the number of non-matching attributes (difference) and $r$ is the number of relevant attributes for the query. This probability is derived using a combinatorial argument, assuming that all attributes are equally likely to be relevant.
They then extend this model by allowing the number of relevant attributes, $r$, to vary according to a probability distribution, $q(r)$. This leads to a probability of successful analogy, $p(d)$, which is a function of the difference $d$ alone. The paper explores different assumptions for $q(r)$ (e.g., uniform, exponential, normal distributions) and shows that the resulting $p(d)$ curves exhibit a consistent shape, closely resembling empirical data.
Datasets
Not discussed in the paper. The paper relies on theoretical analysis and empirical data from existing psychological studies, specifically referencing Shepard's work on stimulus generalization.
Key Insights and Ideas
- Justification for Similarity Heuristic: The paper provides the first quantitative justification for the "maximal similarity" heuristic in analogy, showing that under reasonable assumptions, a more similar source is indeed more likely to lead to a correct analogy.
- Probability of Analogy: A formal framework is presented to calculate the probability of a successful analogy as a function of the difference in attributes between the source and target, and the number of relevant attributes.
- Empirical Alignment: The theoretical predictions of the model closely match empirical data from psychological experiments on stimulus generalization, suggesting that human and animal analogical reasoning might be grounded in similar probabilistic principles.
- Representation Matters: The authors acknowledge that the utility of the simple model relies on the assumption that attributes appropriately "carve up the world" into relevant causal aspects. This highlights the importance of well-structured representations for effective analogical reasoning.
- Generalization of Similarity: The paper touches upon the challenge of defining a representation-independent similarity measure and suggests that the concept of "mini-attributes" and logarithmic scaling might be avenues for future work.
Key Results
The paper demonstrates that the probability of a successful analogy, $p(d)$, decreases as the difference $d$ between the source and target increases. This decrease is generally monotonic and concave. Different assumptions about the distribution of relevant attributes ($q(r)$) lead to similar functional forms for $p(d)$, often approximating an exponential decay. Figure 1 visually presents $p(d, r)$ for various values of $r$, showing how the probability of correctness narrows as the number of relevant attributes increases. Figure 2 illustrates $p(d)$ under different $q(r)$ assumptions, all showing a similar decay pattern. The authors highlight that these theoretical results align remarkably well with empirical stimulus generalization gradients observed by Shepard across a wide variety of domains (Figure 3).
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Representation Dependence: The simple model relies on attribute-value representations and the assumption that attributes are equally likely to be relevant. Generalizing this to arbitrary predicate calculus formulae and handling non-uniform relevance is a significant challenge.
- Similarity Measure Definition: Developing a representation-independent similarity measure that accounts for inexact matches and varying attribute significance is an open problem.
- Integration with Determination: The paper suggests integrating similarity-based reasoning with determination-based reasoning (where relevance is explicitly defined) to create a more robust analogical system. This could involve using overconstrained determinations to circumscribe relevant features, probabilistic determinations to weight attribute contributions, or using high similarity to discover new determinations.
- AI Database Application: The authors plan to explore the implications of their theory by performing large-scale analogies in an AI database (Lenat's CYC system).
- Theory of Natural Kinds: The need for a theory of "natural kinds" for features, analogous to natural kinds for objects, is mentioned as crucial for effective representation.
Summary updated