Home / Curated / Architectures / Communication Neural Network

Communication Neural Network

Architectures AIMech:CommunicationNeuralNetwork DRAFT

Aliases: CommNet, Communication Neural Net, CommNets

A neural architecture for cooperative multi-agent tasks that alternates shared per-agent hidden-state updates with continuous communication formed by averaging other agents’ hidden states, allowing task losses to train communication and agent computation together.

Identity Notes & Concept Scope Compared all 442 pinned AIO and 598 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 8, 0, 0 lines. All distinct match contents were inspected. Semantic matching covered 24 AIO and 51 native definitions, with every matched definition read. 4 full annotated AIO node/edge/generated contexts and 5 full native records were inspected. CommNet is the named Communication Neural Net with shared agent modules and exclude-self mean communication. It differs from VAIN learned-distance weights, Interaction Network pair-specific learned effects, and the maintained GAT, GIN and GraphSAGE constructions. The neural architecture maps broadly to ANN; recurrence and local connectivity are extensions, not essential identity conditions.

Claims & Grounded Evidence

6 Claims

Mechanism #continuous-exclude-self-communication

Encode each agent state-view into an initial hidden vector. At communication step i, a module shared across agents updates each hidden state from its own previous state and communication vector; the next communication vector is the arithmetic mean of the other agents’ updated hidden states. The default initial communication is zero. After K steps, the base controller decodes per-agent action distributions and samples actions. Parameters may differ between communication steps; sharing across agents does not require sharing across steps.

supports Source: Learning Multiagent Communication with Backpropagation · Locator: Section 2.1, PDF p. 2, Equations 1–2 and encoder/decoder description; Figure 1, p. 3

Limitation #mean-and-empty-neighborhood-boundaries

The broadcast mean excludes self and divides by J-1. It differs from a sum and from a mean including self; different multisets can produce the same mean. The displayed formula is undefined for J=1 without an additional convention. The local-connectivity extension averages only over N(j) and is likewise undefined for an empty neighborhood as written. No author implementation was inspected to establish its handling of these cases.

supports Source: Learning Multiagent Communication with Backpropagation · Locator: Section 2.1, Equation 2, PDF p. 2; Section 2.2, Equation 3, p. 3; exact constructed mean examples

Scope #permutation-and-computation-boundaries

With shared agent modules, a consistent permutation of agent inputs permutes the corresponding per-agent output distributions. This is equivariance, not an unchanged ordered output array; the paper describes its block-structured communication matrix as permutation invariant. All broadcast means can be computed by one shared total followed by a per-agent subtraction and division, using linear aggregation arithmetic for fixed feature width. This algebra does not establish physical communication cost, total distributed-system latency or implementation behavior.

supports Source: Learning Multiagent Communication with Backpropagation · Locator: Section 2.1, Equation 2 and block matrix T, PDF p. 2; Figure 1, p. 3; direct permutation and arithmetic-count examples

Scope #extensions-and-training-boundaries

Local connectivity replaces the broadcast mean with a neighborhood mean; skip connections supply the initial agent encoding to later modules. A temporal variant shares the module across time and may use recurrent or LSTM modules, but the basic finite-step feedforward construction need not reach a recurrent fixed point. Continuous inter-agent vectors allow backpropagation through communication; supervised losses or an outer reinforcement-learning procedure supply training signals. Discrete environment-action sampling is not thereby made directly differentiable, and the discrete-symbol baseline is a separate construction.

supports Source: Learning Multiagent Communication with Backpropagation · Locator: Sections 1–2.2, PDF pp. 1–3, Equations 3–4; Related Work and discrete baseline, p. 4, Equations 5–6

Scope #cooperative-task-evaluation

The paper assumes fully cooperating agents with a shared reward. It evaluates lever assignment, simulated traffic and combat, and a sentence-agent question-answering variant. The lever experiment selects five agents from a pool of five hundred and reports distinct-lever coverage, while traffic reports failure rates and combat reports win rates for specified MLP/RNN/LSTM variants. These task-specific metrics and selected seeds do not demonstrate general coordination, safety or superiority in arbitrary environments.

supports Source: Learning Multiagent Communication with Backpropagation · Locator: Introduction, PDF p. 1; Sections 4.2–4.3, pp. 4–7, Tables 1–3; Section 4.4, p. 8

Limitation #global-answer-variant-and-evidence-limits

The bAbI variant treats story sentences as agents, sums final hidden states after two communication steps and uses a global answer decoder instead of the base per-agent action heads. On the reported 10K tasks, Table 4 gives CommNet mean error 7.1%, compared with 4.2% for MemN2N and 2.8% for DMN+, so the results do not support universal superiority. Communication-vector PCA and braking associations are illustrative observations, not general causal explanations. Very large populations and heterogeneous agent types are left for further work.

supports Source: Learning Multiagent Communication with Backpropagation · Locator: Section 4.3.2 and Figure 3, PDF pp. 6–7; Sections 4.4–5 and Table 4, p. 8

⚠️ Curated Limitations & Negative Results

The named identity is Sukhbaatar, Szlam and Fergus’s Communication Neural Net (CommNet), not every communication network or language-model agent system.
The broadcast equations assume at least two agents; the local formula requires a nonempty chosen neighborhood. No author-code audit supplies omitted edge-case conventions.
Per-agent output distributions are permutation equivariant under consistent relabeling. The global question-answering variant has a different final readout; the two are not interchangeable.
All nine main-paper pages were read. Supplementary material, author code and other versions were not inspected. Tiny exact calculations check pooling identities only; no training, hardware experiment or benchmark replication was performed. DRAFT remains.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	broad_match	Shared learned per-agent modules, continuous hidden communication and output decoders form a particular artificial neural architecture. The target is broader; the basic architecture does not require temporal recurrence, the named GCN operator, a probabilistic graphical-model task or learned attention weights.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`sukhbaatar2016`	Learning Multiagent Communication with Backpropagation Sainbayar Sukhbaatar, Arthur Szlam, Rob Fergus	primary paper full_text	ID: https://papers.neurips.cc/paper_files/paper/2016/hash/55b1927fdafef39c48e5b73b5d61ea60-Abstract.html sha256:db09430a0aeb…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-13T09:30:19Z

Updated: 2026-10-13T09:30:19Z

Source File: data/curated/CommunicationNeuralNetwork.yaml