AI Glossary / Graphs and symbolic AI / Intermediate

Social laws

Definition

In multiagent AI, social laws are shared rules or conventions that constrain agents' behavior so that they can operate together more predictably. They specify which actions are permitted in relevant circumstances, reducing conflicts among independently acting participants. Here, law is a technical coordination term, not a claim about legislation or a system's legal authority.

What it means

One agent's sensible action can interfere with another agent's equally sensible action. Two robots may both try to enter a narrow passage; two software agents may both attempt to reserve the same resource. A shared convention can settle some of these interactions before each individual conflict needs negotiation.
The rules need not prescribe every action. They can leave each agent free to plan its own task within agreed constraints. The useful question is whether those constraints provide the desired joint behavior while leaving enough freedom for agents to accomplish their goals.

How it works

Describe the agents, relevant shared state, available actions, and possible conflicts. Express the convention as conditions governing allowable actions. Analyze the resulting joint behavior, including simultaneous choices and failure cases.
For example, an entry rule may depend on a reservation state, not merely on whether a doorway currently looks empty. Agents must apply compatible interpretations of that state. Check both safety, such as avoiding conflicting occupancy, and progress, such as allowing a waiting participant eventually to proceed. A rule that avoids all movement may prevent collisions while failing every transportation goal.

Three practical examples

• A robot simulation can require a shared token before entering a passage that cannot support simultaneous traffic.
• A collection of task agents can follow a reservation protocol so that only one agent changes a shared item at a time.
• A multiagent planning exercise can allocate compatible roles before a joint activity, reducing repeated disagreement over who performs which action.

Think of it this way

Think of a shared workshop where everyone agrees to sign out a tool before using it. The convention helps people plan around one another without negotiating each use from scratch. The analogy is limited because software agents need precise rules for simultaneous requests, stale information, crashes, and release of the tool; informal human understanding is insufficient.

Worked example

Consider an original two-agent model with agents A and B and one shared resource. Each agent can choose Enter or Wait at a decision step. If both enter when the resource is free, the model records a conflict.
Without a coordinating rule, the four joint choices are Enter/Enter, Enter/Wait, Wait/Enter, and Wait/Wait. One is conflicting, but this count does not imply a 25% conflict probability unless those four choices are equally likely.
Introduce a token variable T that takes values A or B. Initially T = A. The social law permits Enter only when the resource is free and the entering agent holds the token. When T = A and the resource is free, B must wait; A may enter or wait. Thus the simultaneous Enter/Enter choice is no longer permitted.
Suppose A enters at step 0, uses the resource for one step, and releases it while passing the token to B at step 1. B can then enter at step 1. Under these stipulated transition rules and compliant execution, no step admits two simultaneous entrants.
However, if A keeps choosing Wait and the protocol has no token-transfer rule for that case, B may wait forever despite the resource being free. The collision-avoidance argument has not established progress. Adding a precise idle-transfer rule changes that behavior; crash recovery and simultaneous state updates would still need their own model.

Illustrative case study

In a fictional delivery-robot laboratory, a team first adopts the instruction to enter a doorway only when it is empty. A simulator reveals that two robots can observe the same empty doorway and move toward it simultaneously.
The team replaces that instruction with an explicit reservation state and tests transitions involving release, cancellation, and waiting. It also checks whether one robot can monopolize access. The case illustrates why a convention must be analyzed as a joint process. It does not report a real deployment, a measured reduction in collisions, or a guarantee about physical hardware.

Common misconceptions and mistakes

Shared intentions are not the same as compatible action rules. A rule can be ambiguous even when every agent is trying to cooperate. Observing a resource as free is not necessarily equivalent to obtaining an exclusive reservation.
Do not infer fairness from mutual exclusion, or eventual completion from the absence of collisions. More restrictive laws are not automatically better: prohibiting all actions removes useful behavior. Conversely, a rule that works under perfect compliance and synchronized state updates may fail when agents have delayed information or different interpretations.

Business use

The concept is useful when multiple automated participants share resources or responsibilities. It encourages designers to make interaction assumptions explicit rather than relying on each component to behave reasonably in isolation.
For a proposed coordination rule, record the permitted actions, the state used to decide permission, and how that state changes. Review a few adverse joint traces before scaling up. Evaluate waiting time, task completion, and resource utilization separately, because a single desirable property cannot stand in for all operational goals.

People also ask

Are social laws necessarily learned?

No. They can be designed in advance, synthesized from a formal model, or studied as conventions; learning is not part of the definition.

Do they eliminate the need for communication?

Not necessarily. Some rules reduce negotiation, while others require messages to establish reservations or shared state.

Are they the same as a centralized joint plan?

No. They can constrain independently chosen plans without specifying every participant's complete sequence of actions.

Check your understanding

A rule allows only the token holder to enter a resource. Does this alone prove that every waiting agent will eventually enter?

Answer: No. It can establish exclusive entry under stated assumptions, but eventual access also depends on token transfer, termination, scheduling, and failure behavior.

Related terms

• Multiagent systems
• Joint action
• Multiagent system
• Coordination
• Convention
• Mutual exclusion
• Deadlock
• Fairness
• Distributed planning

Learn next

Next, write the token example as a small state-transition diagram. Add an idle-transfer rule and trace what happens when A waits twice. Then distinguish a collision, a deadlock, and starvation: simultaneous incompatible use, inability to make progress, and one participant being postponed indefinitely are different failure patterns.

More in Graphs and symbolic AI

• Small-scale learning - Small-scale learning refers to machine learning with a relatively limited training dataset.
• Small-world - A small-world network is a graph that combines strong local clustering with relatively short routes between nodes across the network.
• SMDP - SMDP stands for semi-Markov decision process: a model for sequential decisions in which the time between decision points can vary.
• Soft margin - A soft margin is the allowance for penalized margin violations in a support vector classifier.
• Software architecture - Software architecture is the organization of a software system's major parts, their responsibilities, and the ways they interact.
• Spanning Trees - A spanning tree of a connected, undirected graph is a selection of its edges that connects every vertex without creating a cycle.

Sources

Shoham and Tennenholtz, On the Synthesis of Useful Social Laws for Artificial Agent Societies: https://cdn.aaai.org/AAAI/1992/AAAI92-043.pdf AIMA original glossary sentences, coordination conventions: https://raw.githubusercontent.com/aimacode/aima-glossary/master/sentences.txt

Previous: Social data science Next: Social media mining