Convention (Lewis)

David Lewis’s (1969) game-theoretic analysis of conventions: a regularity R in the behaviour of a population is a convention iff almost everyone conforms to R, almost everyone expects almost everyone else to conform, almost everyone prefers conforming on the assumption that others conform, and some alternative regularity could equally have served. Sustained by Mutual Expectation / common knowledge. Foundational analysis of how shared meaning, currency, and social roles arise without explicit agreement.

In this vault

Backlinks

Linked Pages

Language Games for Autonomous Robots

Reference: Steels, L. (2001). IEEE Intelligent Systems, Sept/Oct 2001. Source file: steels01languageGames.pdf. URL

Summary

Steels presents language games as a paradigm for creating grounded, open-ended dialogue between autonomous robots and humans. A language game is a scripted verbal interaction situated in a shared environment that simultaneously exercises perception, conceptualization, verbalization, interpretation, and action. By playing repeated games, agents integrate disparate AI components (vision, speech, planning, learning) into a coherent whole and bootstrap shared vocabulary and meaning without a central designer.

The paper details the guessing game as a canonical example: a speaker picks a topic, conceptualizes a distinguishing feature, verbalizes it, and the listener tries to identify the referent — feedback updates their lexicons’ association scores. Experiments on Sony AIBO and the “Talking Heads” setup show how robots acquire words like color and spatial terms through embodied interaction. Language games thus provide a framework for studying the emergence of grounded communication conventions.

Key Ideas

Language games integrate perception, conceptualization, and action.
Guessing game as a minimal grounded-dialogue protocol.
Shared context + feedback drives lexicon convergence.
Talking Heads experiment: lexicons emerge via population interaction.
Language games tackle integration and grounding jointly.

Connections

Conceptual Contribution

Claim: Grounded, open-ended communication in robots can emerge from repeated situated “language games” that co-exercise perception, conceptualisation, verbalisation and action — with no central designer of the lexicon.
Mechanism: Steels formalises the guessing-game protocol: speaker selects a topic in shared scene, computes a distinguishing category, encodes it as a word; listener decodes, points, and success/failure feeds back to associative lexicon scores. Embodied populations of Sony AIBOs and the Talking Heads platform empirically show lexicon convergence over colour, spatial and proper-name vocabularies, demonstrating that integration of AI sub-systems and semantic grounding are solved jointly rather than separately.
Concepts introduced/used: Language Game, Guessing Game, Grounding, Lexicon Convergence, Talking Heads Experiment, Semiotic Cycle, Emergent Communication
Stance: empirical
Relates to: Provides an embodied-situated counterpart to the decision-theoretic language emergence of Towards Automating the Evolution of Linguistic Competence and the neural-MARL emergence of Emergence of Grounded Compositional Language in Multi-Agent Populations; its feedback-driven convergence contrasts with the top-down protocol stipulation of FIPA-ACL.

In this vault

Mutual Expectation

The infinite hierarchy of expectations that sustains a Convention (Lewis): each party expects others to follow R, expects others to expect them to follow R, and so on. Equivalent to (and rediscovered as) common knowledge in game theory and distributed systems. The operative epistemic condition that makes a regularity self-sustaining without explicit agreement.

In this vault

Common Knowledge (Convention)

The infinite hierarchy of mutual expectation introduced by Lewis (1969) and rediscovered by Aumann (1976) in game theory and Halpern–Moses (1990) in distributed systems: A knows p, A knows that B knows p, A knows that B knows that A knows p, … all the way up. A convention is sustained by common knowledge of the regularity. The construction has the same mathematical content in three independent fields — Lewis-style social philosophy, game theory, and distributed-systems epistemology — and is unattainable in fully asynchronous distributed systems (Halpern & Moses), with deep consequences for what coordination protocols can achieve.

In this vault

Coordination Game

In game theory, a strategic situation with multiple Pareto-comparable equilibria where players prefer reaching any equilibrium to none, with limited or no preference over which one. The driving-side game (drive-on-the-right vs drive-on-the-left), Schelling’s meeting-place games, and the conventional-language game are paradigm examples. Lewis’s conventions are precisely the regularities that solve coordination games for a population of agents.

In this vault

Convention - Lewis

Convention: A Philosophical Study

Reference: Lewis, D. K. (1969). Convention: A Philosophical Study. Harvard University Press, Cambridge MA. (Reprinted 2002, Blackwell.) SEP entry on Convention · Wiley publisher page (book; no full-text OA.)

Summary

Lewis offers the first rigorous game-theoretic analysis of how shared conventions — including, paradigmatically, the conventions of language — arise and sustain themselves among rational agents without explicit agreement. The book opens by defining a coordination problem: a situation in which several agents have multiple equilibria available, all parties prefer reaching some equilibrium to none, and they are largely indifferent (or have non-conflicting preferences) over which equilibrium they reach. Lewis’s central thesis is that a convention is a regularity in behaviour that is one solution to a coordination problem, sustained by mutual expectation — each party expects the others to follow the regularity, expects the others to expect them to follow it, and so on up an infinite hierarchy that Lewis was the first to formalise as common knowledge. The book applies this analysis to driving conventions (drive on the right vs left), signalling, currency, and at length to language: the conventions of meaning are coordination equilibria sustained by mutual expectation in a population of speakers; semantic content is what would be the case if the convention were followed. Lewis’s framework supplies the philosophical foundation for every later treatment of emergent and negotiated meaning — emergent-communication research in MAS and deep learning, the ACL evolution programme, and the language-games tradition (Wittgenstein operationally, Steels formally) all proceed in the conceptual space Lewis defined. The common-knowledge construction is independently a foundational result for distributed-systems theory (Halpern & Moses 1990) and game theory (Aumann 1976).

Key Ideas

Coordination problem: a strategic situation with multiple Pareto-comparable equilibria where no party prefers a non-equilibrium outcome and the parties’ preferences over equilibria largely coincide. The classical examples — driving side, meeting place — generalise to language, money, social roles.
Convention as a self-sustaining regularity: a regularity R is a convention iff (i) almost everyone conforms to R, (ii) almost everyone expects almost everyone else to conform to R, (iii) almost everyone prefers conforming on the assumption that others conform, and (iv) some alternative regularity could have been the convention instead (no party of the convention is intrinsically preferred).
Common knowledge: the infinite hierarchy of mutual expectation that sustains a convention — A knows that B follows R; A knows that B knows that A follows R; and so on. Lewis was the first to give a formal account of this in social terms; the construction was rediscovered by Aumann in game theory and Halpern–Moses in distributed systems.
Salience and precedent as convention selectors: when multiple equilibria are available, salient features (Schelling’s focal points, see The Strategy of Conflict) and historical precedent select among them. There is no rationality-driven argument for which equilibrium becomes conventional; the answer is contingent on saliency.
Language as convention: meaning is what would obtain if the convention were followed. Truth-conditions, communicative force, and reference are all conventionally fixed — but the convention is not arbitrary, it is sustained by everyone’s expectation that everyone else will use words a particular way.
Conventions vs explicit agreements: Lewis argues that explicit agreement is neither necessary nor sufficient for convention — many conventions arise without agreement (precedent does the work) and many agreements never become conventions (because mutual expectation does not develop).
Population vs pair: conventions are population-level facts; one cannot have a convention with a single partner. Mutual expectation must extend across the relevant community for the regularity to be convention-grade.

Connections

Conceptual Contribution

Claim: Conventions — including the conventions of language — are best analysed as solutions to coordination problems sustained by mutual expectation (formally, common knowledge) in a population of rational agents. There is nothing arbitrary about the convention’s function, but the particular regularity that becomes conventional is contingent on salience, precedent, and historical accident.
Mechanism: Game-theoretic analysis of coordination problems with multiple equilibria; definition of convention via four clauses (conformity, mutual expectation, conditional preference, alternativeness); explicit construction of common knowledge as an infinite hierarchy of expectations; application to driving conventions, signalling, currency, and at length to language semantics.
Concepts introduced/used: Coordination Game, Convention (Lewis), Common Knowledge (Convention), Mutual Expectation, Focal Point (via dialogue with Schelling), Salience, Precedent, Convention vs Agreement.
Stance: foundational philosophical monograph.
Relates to: Direct philosophical foundation for every later treatment of emergent meaning and negotiated language: Steels’ language games are operational realisations of Lewis’s coordination-problem analysis with explicit interaction protocols; Lazaridou et al. and the deep-learning emergent-communication line train neural agents to converge on Lewis-style conventions in referential games. The ACL-evolution programme (Toward Automated Evolution of ACLs, Towards Automating the Evolution of Linguistic Competence) treats the ACL itself as an evolving Lewis-convention. Lewis’s common-knowledge construction is independently the foundation of Halpern & Moses 1990 in distributed systems and Aumann (1976) in game theory — distributed protocols and conventional language share the same mathematical object. The framework also clarifies what is at stake in the Mentalistic Semantics vs Public Semantics debate in ACLs: Lewis-style conventions are public-semantics objects (mutual-expectation regularities) but their stability depends on something close to mentalistic conditions (each agent’s expectations about others’ beliefs), so the two semantic families are dual aspects of the same phenomenon rather than competitors. Why AI Agents Communicate In Human Language revisits Lewis from the LLM-agent angle: natural language is a particular Lewis-convention (a contingent equilibrium) and not necessarily the optimal one for inter-agent coordination.