Focal Point

Schelling’s (1960) concept: in a coordination game with multiple equilibria, the equilibrium parties converge on by virtue of salience — some feature that “stands out” in their shared informational and cultural background, distinguishing it from the others. Salience is not derivable from the payoff structure alone — it depends on context, prominence, simplicity, history. The cognitive mechanism by which Lewis-style conventions emerge without explicit agreement.

In this vault

Backlinks

Linked Pages

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.

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

Tacit Coordination

Coordination achieved without explicit communication — parties select compatible actions by relying on shared salience (Focal Point), shared convention, or shared background knowledge. Schelling’s experiments (1960) demonstrated robust tacit coordination in strangers (e.g. converging on Grand Central at noon). In MAS the concept underlies zero-shot coordination, ad-hoc teamwork, and emergent role assignment without negotiation; in LLM-agent settings, it is the principal mechanism for coordination across agents that share no explicit protocol.

In this vault

Schelling Point

Synonym for Focal Point — the salient equilibrium parties converge on in a coordination game with multiple equilibria, named after Thomas Schelling who introduced the concept in The Strategy of Conflict (1960).

In this vault

The Strategy of Conflict

Reference: Schelling, T. C. (1960). The Strategy of Conflict. Harvard University Press, Cambridge MA. (Reprinted 1980 with new preface.) Harvard UP page · Internet Archive borrow (book; no full-text OA.)

Summary

Schelling reframes game theory as the theory of strategic interdependence — the analysis of decision-making in situations where each party’s best move depends on what they expect the others to do, including in situations of mixed conflict and cooperation. The book opposes the prevailing post-von-Neumann emphasis on zero-sum games and pure conflict; Schelling’s central observation is that almost all real strategic situations have both conflict and cooperation elements, and the interesting analytic work is in the cooperation-within-conflict structure. The book’s most enduring concept is the focal point (or Schelling point): in a coordination problem with multiple equilibria, parties converge on the equilibrium that is salient — distinguished by some feature that “stands out” in the shared cultural / informational background. Schelling demonstrated this experimentally: subjects asked to meet a stranger in New York City on a particular day, with no prior communication, overwhelmingly converged on Grand Central Terminal at noon. There is no equilibrium-selection argument in classical game theory that picks Grand Central; saliency does. Other lasting contributions include the analysis of credible commitment (binding oneself in advance can be a strategic asset, not a constraint); the concept of the threat that leaves something to chance; precommitment via burning bridges; tacit bargaining; and an early treatment of arms-race dynamics. The book inaugurated the modern study of bargaining, conflict resolution, and coordination — and won Schelling the 2005 Nobel Prize in Economics. For agent communication, the central inheritance is the focal-point concept: when explicit communication is impossible or noisy, agents coordinate on salient features of their shared situation. Lewis’s <em>Convention</em> (1969) is its philosophical companion.

Key Ideas

Strategy is interdependent decision: the right object of study is the situation in which each party’s payoff depends on others’ actions, not the isolated decision-theoretic agent.
Mixed-motive games: real conflicts almost always combine conflicting and shared interests; pure-conflict zero-sum games are a degenerate special case.
Focal points / Schelling points: in coordination problems with multiple equilibria, parties converge on the equilibrium that is salient in the shared informational and cultural background. Salience is not derivable from the payoff structure alone — it depends on context, history, prominence, simplicity.
Credible commitment as strategic asset: a party that can publicly bind itself in advance often does better than one with full strategic flexibility, because the binding rules out outcomes the other party would prefer to avoid. Burning bridges, hostage-taking, automatic-response systems, and unilateral declarations are concrete devices.
The threat that leaves something to chance: an irrevocably probabilistic threat (a “doomsday machine” with random triggering, a brinksmanship strategy) can credibly deter when a deterministic threat could not — because the threatener does not control whether it will be carried out.
Tacit bargaining: parties can coordinate on outcomes without explicit communication by relying on shared salience and convention.
Arms-race stability vs instability: Schelling distinguishes situations in which arms accumulation reduces incentives to strike first (stabilising) from those in which it increases them (destabilising) — the foundational analysis behind nuclear deterrence theory.

Connections

Conceptual Contribution

Claim: Strategy is the analysis of interdependent decision-making in mixed-motive (conflict + cooperation) situations; the right concepts include credible commitment, focal-point convergence under saliency, the threat that leaves something to chance, and tacit bargaining. These are not derivable from classical zero-sum game theory.
Mechanism: Reframing of game theory as theory of strategic interdependence; experimental and conceptual demonstrations of focal-point coordination; analysis of commitment devices and probabilistic threats; case studies in industrial bargaining, nuclear deterrence, racial integration, and traffic.
Concepts introduced/used: Focal Point, Schelling Point, Tacit Coordination, Credible Commitment, Threat Leaving Something to Chance, Mixed-Motive Game.
Stance: foundational research monograph (Nobel-Prize-winning).
Relates to: Direct philosophical and analytical companion to Lewis 1969 — both treat the equilibrium-selection problem in coordination games, with Lewis emphasising the social-philosophical structure of mutual expectation and Schelling emphasising the cognitive structure of saliency. The two together supply the foundation for every analysis of how shared meaning, currency, social roles, and (now) ACL semantics arise without explicit agreement. In MAS, the focal-point concept underlies role assignment without communication, emergent coordination conventions in ad-hoc teamwork, and the whole study of zero-shot coordination in MARL. In the ACL evolution programme (Toward Automated Evolution of ACLs, Towards Automating the Evolution of Linguistic Competence), focal points appear implicitly as the salient features that drive lexicon convergence. In LLM-agent settings, focal-point reasoning is one of the principal mechanisms by which agents coordinate without protocols — and a primary attack surface, since what is salient to an LLM is shaped by training data in adversarially exploitable ways.