An Agent-Based Model Studying the Acquisition of a Language System of Logical Constructions

Josefina Sierra-SantibáñezAAAI 2014 · DOI

Introduction

How can a group of individuals with no prior shared language agree on a way to communicate complex logical concepts? This fundamental question in evolutionary linguistics was explored in a paper presented at AAAI 2014. While previous research has demonstrated how agents can develop shared vocabularies for physical objects or spatial relations, the emergence of logical constructions—such as AND, OR, and NOT—presents a significantly higher level of complexity.

The Challenge of Abstract Logic

Logical functions like "exclusive or" (XOR) or "implication" (IF-THEN) are abstract; they lack direct physical referents in the world. To communicate these concepts effectively, a population of agents must reach a consensus on two critical fronts:

  1. Vocabulary: Assigning arbitrary strings to represent specific logical functions.
  2. Grammar (Word Order): Establishing syntactic conventions to determine the relationship between arguments (e.g., distinguishing "A implies B" from "B implies A").

The study aims to demonstrate that these complex systems can emerge through self-organization, driven by social interaction rather than pre-programmed linguistic rules or central authority.

The Agent-Based Model

The researcher implemented an Agent-Based Model (ABM) in Prolog, where autonomous software agents interact through "language games."

Language Games

In each interaction, agents are paired as a "Speaker" and a "Hearer." The speaker attempts to describe a subset of objects using a logical formula (e.g., [and, up, left]). The game is considered successful if the hearer correctly interprets the utterance. Notably, the model allows for logical equivalence: a game succeeds if the hearer's interpretation is logically equivalent to the speaker's intent, even if their internal representations differ.

Cognitive Mechanisms

The agents rely on three primary mechanisms to evolve their language:

  • Invention and Adoption: Agents invent random strings when they lack a way to express a concept and adopt new strings heard from others.
  • Induction (Simplification & Chunking): Agents look for patterns across different utterances to induce shared structures and logical operators.
  • Adaptation via Lateral Inhibition: Agents track the "usefulness" of grammar rules using a scoring system. When a game succeeds, the score of the utilized rule increases, while the scores of competing rules (synonyms or alternative word orders) decrease.

Evolutionary Dynamics and Cultural Transmission

A key aspect of the study is the "generational setup." To test the stability and transmittability of the emerging language, the population is divided into "elders," "adults," and "young." Every 500 games, the elders are replaced by new, "empty" agents.

Figure 1: Evolution of communicative success and coherence across generations, showing the impact of population turnover every 500 games.

As shown in Figure 1, the population experiences a temporary dip in communicative success and coherence whenever a generation is replaced. However, the system quickly recovers. Interestingly, this "bottleneck effect" helps stabilize the language; new agents force the language to become more regular and easier to learn, as overly idiosyncratic rules fail to persist across generations.

Competition and Convergence

The emergence of a unified grammar is a process of competition between various possible constructions.

Figure 3: The competition between different linguistic constructions over time, where one dominant construction eventually emerges through lateral inhibition.

Figure 3 illustrates how different constructions (represented by their average scores) compete within the population. Through lateral inhibition, the agents eventually converge on specific word orders to disambiguate non-commutative functions like IF-THEN. Once a dominant construction is established, it is transmitted effectively to subsequent generations, maintaining high communicative success (above 0.96) despite total population turnover.

Conclusion

The study demonstrates that compositional grammars for logical constructions can emerge and stabilize through decentralized social interaction. While the current model uses a small population and non-recursive formulas, it provides a robust framework for understanding the cultural evolution of language. Future work will look toward scaling these dynamics and testing them in embodied systems, such as physical robots, to account for real-world sensory uncertainty.

Summary updated

BibTeX
@inproceedings{sierrasantibanez2014agent,
  author = {Josefina Sierra-Santibáñez},
  title = {{An Agent-Based Model Studying the Acquisition of a Language System of Logical Constructions}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.8749},
}
Figure