Economic Attention Networks

help edit space_dashboard

Economic Attention Networks: Associative Memory and Resource Allocation for General Intelligence

Authors: Matthew Iklé, Joel Pitt, Ben Goertzel, George Sellman
Year: 2009
Venue: Proceedings of AGI-09, Atlantis Press
Links: PDF

Summary

Introduces ECAN — Economic Attention Networks — which use economic metaphors (currency, markets) to allocate attention across a knowledge hypergraph.

Relevance to Hyperon

ECAN is a core Hyperon component. This paper defines the attention allocation mechanism that determines which atoms in the metagraph receive processing resources.

Curated Excerpts (cluster-pilot extraction, 2026-04-25; updated 2026-04-26 with ECAN cluster Sources 3+4)

Curated notes, excerpts, and equations from the full source text at publication_texts/json/Economic_Attention_Networks.json. See RawData+Publications+Economic Attention Networks for the complete raw extraction.

STI/LTI as Conserved Currency

STI and LTI are explicitly treated as artificial currencies subject to conservation laws (modulo controller-driven inflation/deflation). This is mathematically rigorous resource allocation, not metaphor: rent extracts STI from low-attention atoms; wages distribute STI to atoms participating in valued activity; an OpenCog-Prime "central bank" mediates conversion rates.

Key Update Equations

  • Rent: rent = Rent * max(0, log2(si / recentMaxSTI))^2
  • Wages (positive stimulus): wages = WageStimulus * (sum(pi) / n)
  • Hebbian update: c_ij' = ConjDecay * (si * sj / recentMaxSTI^2) + (1 - (si * sj / recentMaxSTI^2)) * c_ij
  • Diffusion (stochastic decision): P(diffuse) = 0.5 * (1 + tanh(shape * (si - FocusBoundary)))

Attentional Focus + Hebbian/InverseHebbian Links

Atoms above an STI threshold form the Attentional Focus. Hebbian probability values represent odds that if one atom is in focus, another also is; InverseHebbian values represent odds of exclusion.

Convergence Result + Variant 2

The paper proves convergence for a subset of hyperbolic decision functions via a contraction-mapping argument. A separate Variant 2 is designed around AVDIFF as a Lyapunov-style function. Associative-memory experiments show capacity comparable to a Hopfield net only when ECAN parameters allow imprinted memories to fit into the attentional focus — the paper itself flags this caveat.

ECAN as PLN Inference Control

The paper proposes ECAN to guide PLN forward/backward chaining by weighting possible inference steps using STI values of involved atoms; inference results would reward atoms with STI/LTI according to utility. This proposal was realized only narrowly in OpenCog Classic, then deliberately removed. ECAN cluster pilot Source 3 (2026-04-26) established the executable lifecycle:

  • Wired up in URE forward chaining: commit 0a0b09912 (2016-03-09 Misgana Bayetta, "Pass source selection mode to the constructor") added a source_selection_mode enum with an STI case using tournament_elem[s] = s->getSTI() for tournament selection of inference sources.
  • Unwired: commit 0b744dbab (2018-10-23 Nil Geisweiller, "Re-work source and rule selection in the forward chainer") commented out the STI case with the explicit comment "An attentionbank is needed in order to get the STI...". The 2018 rewrite authors knew the AttentionBank dependency was required and chose not to wire the new SourceSet path through it.
  • Constructor argument removed: commit 5a5b7785d (2019-01-08 Nil Geisweiller). The enum value at ForwardChainer.h:42 persists as dead residue at current URE HEAD.

The hook read atom STI for source selection only — it was not a full ECAN agent integration and did not couple PLN reward signals back to ECAN currency dynamics. URE subsequently invested in Bayesian Thompson sampling, MixtureModel, BIT, and trace/control-rule meta-learning as its durable inference-control architecture (a Bayesian/control-rule route, decoupled from any global AttentionBank). Modern Hyperon-MeTTa PLN (trueagi-io/PLN's lib_pln.metta) has no attention-layer integration; the chaining/ MeTTa successor preserves only ~5/16 URE inference-control features.

Implementation Status

ECAN attention-currency dynamics are implemented in OpenCog Classic attention(C++) and ported to MeTTa as metta-attention.0/4againsttanh Both implementations score strict-literal compliance the 2009 paper's update equations: Classic C++ already used pragmatic engineered surrogates (target-funds/buffer rent, fund-buffer wage adjustment, normalized-STI fuzzy-conjunction Hebbian, deterministic elapsed-time decay in place of the literal sigmoid), and the MeTTa port faithfully ports those C++ surrogates rather than implementing the paper math. The MeTTa port did not lose paper equations — Classic C++ had already moved away from them.

Beyond the narrow URE STI source-selection hook (2016–2018), the ECAN-PLN runtime control loop the paper describes is paper-architectural for the current Hyperon-MeTTa cluster. Production PLN/, PLN3, and the CeTTa PLN line all run on confidence-priority forward-chaining queues with no attention-layer integration. Re-establishing executable PLN-attention coupling is an open problem; see ECAN Full → Development and Historical Context.