Home / Curated / Architectures / Guarded discrete stochastic network

Guarded discrete stochastic network

Architectures AIMech:GuardedDiscreteStochasticNetwork DRAFT

A binary neural constraint-satisfaction architecture whose main assignment neurons are asymmetrically coupled to empty-row guards and updated by stochastic row selection and maximal inconsistency.

Identity Notes & Concept Scope The actual 2026-10-15 post-cohort 350 collection at commit f654ed4db7fe7277efa4478aa4e71eddd8085d97 covers 781 maintained and 442 generated records. Four fixed-word searches included ignored and hidden files across the original checkout and three retained discovery/recovery checkouts. All 20338 occurrences were reconstructed into 1167 distinct contents, freshly read by two attributed supporting readers. Root read 29 focused contents, all six normalized identity/definition rows and eleven contextual fields, every matching field and owner projection for 24 frontier objects, and both selected owners in full. Normalization covered 4861 identity-field occurrences. No candidate collision was found in this bounded scope. This source-scoped architecture is grounded in the 1989 guarded neural CSP construction and preserves the separately unread 1990 comparison. The supplied guards, asymmetric coupling and row-wise maximal-inconsistency updates distinguish it from GENET learning and symmetric Hopfield memory. The maintained Min-conflicts heuristic is a symbolic approximation context, not the same neural state dynamics. The broader AIO ArtificialNeuralNetwork class is explicitly mapped as broad_match; no exact aliases or version equivalence are asserted.

Claims & Grounded Evidence

5 Claims

Mechanism #binary-neural-constraint-encoding

In the inspected binary finite-CSP encoding, a binary neuron y_ij represents assigning value j to variable i. Its main-network input is a bias plus the weighted outputs of other neurons, with activation one at nonnegative input and zero at negative input. A positive bias beta, inhibitory weight -omega for incompatible value pairs and -eta between distinct values in one variable row, with omega and eta greater than beta, encode the pair constraints and at-most-one assignment preference.

supports Source: Learning in Stochastic Neural Networks for Constraint Satisfaction Problems · Locator: Section 2, PDF pp. 2–3 / printed pp. 368–369, Equations (1)–(4) and Figure 1. Equation (2) visually uses input >= 0.

Mechanism #asymmetric-empty-row-guards

An auxiliary guard neuron monitors each variable row. Its total input is gamma-theta*sum_j(y_ij), with gamma>0 and theta>gamma, so it is active exactly when that row is empty. An active guard supplies sufficiently strong positive excitation to the main neurons in its row, overcoming possible inhibition in the finite network. The inspected update convention maintains each guard’s output consistent with the main state on a faster timescale; the combined coupling is asymmetric.

supports Source: Learning in Stochastic Neural Networks for Constraint Satisfaction Problems · Locator: Section 2, PDF pp. 3–4 / printed pp. 369–370, guard bias and connections, Figure 2 and guard-update paragraph. The total-input expression includes the separately stated positive guard bias.

Mechanism #random-row-maximal-inconsistency-update

The main update selects a guarded row at random. If that row has inconsistent neurons, it changes one with maximal inconsistency: the nonnegative input z_ij for an off neuron with z_ij>=0, or |z_ij| for an on neuron with z_ij<0. Ties are broken arbitrarily. This is a row-based selection followed by one main-neuron transition, with guards maintained consistent; no uniform distribution over rows is specified.

supports Source: Learning in Stochastic Neural Networks for Constraint Satisfaction Problems · Locator: Section 2, PDF p. 4 / printed p. 370, Equation (5) and adjacent selection/guard paragraphs. The equality in the off-neuron condition is visually resolved rather than taken from the incomplete text extraction.

Mechanism #initialization-and-virtual-connections

Heuristic assignments can seed the main state; otherwise the source recommends starting with all main neurons off. Connection values need not all be stored in advance: they can be computed when a participating neuron first turns on, and may be recomputed instead of stored when efficient. This is an optional representation strategy, not a universal storage or runtime bound.

supports Source: Learning in Stochastic Neural Networks for Constraint Satisfaction Problems · Locator: Section 2, PDF p. 4 / printed p. 370, initialization and virtual-connection paragraphs.

Scope #optional-learning-and-symbolic-approximation-boundary

The 1989 paper presents the guard architecture separately from a learning module that examines recorded connections from training runs and adjusts biases or connections through partial consistency inference. Such learning is not required to define every GDS execution. The 1992 report describes symbolic min-conflicts hill climbing as an approximation to GDS behavior, rather than asserting identical state trajectories or identifying the neural architecture with that heuristic.

supports Source: Learning in Stochastic Neural Networks for Constraint Satisfaction Problems · Locator: Section 2, PDF pp. 2–4 / printed pp. 368–370; Section 3, PDF p. 6 / printed p. 372, separate learning module and partial-consistency scope; Section 4, PDF p. 9 / printed p. 375.
supports Source: Minimizing Conflicts: A Heuristic Repair Method for Constraint-Satisfaction and Scheduling Problems · Locator: Section 3.2, PDF p. 9 / printed p. 4, symbolic approximation; Section 4.4, PDF pp. 18–19 / printed pp. 13–14, bounded task comparison.

⚠️ Curated Limitations & Negative Results

The operational source is the acquired 1989 NASA conference paper. The exact 1990 IJCNN full text remains unread; no identity of every equation, update choice or experiment across those versions is asserted.
The fixed-weight binary-CSP encoding assumes finite domains, positive bias and stronger inhibitory weights. Each nonempty variable domain needs a guard excitation large enough for the possible inhibition in that finite network; no single arbitrary finite gain is asserted to work for all problem sizes.
Asymmetric coupling to guards removes the ordinary symmetric Hopfield convergence guarantee. A stopping criterion may be needed; stopping without a satisfying state is not an unsatisfiability certificate or proof that no useful fixed point exists.
Exactly one active neuron per variable is a solution condition, not an invariant throughout the dynamics. Empty rows can occur and activate their guards. At-most-one is preserved under the stated one-neuron update and initially at-most-one state, but is not asserted for arbitrary initialization.
The source specifies random row selection and arbitrary ties without fixing a unique random distribution, tie policy or universal parallel schedule. One main-neuron transition with immediately consistent guards is not a synchronous update of every main neuron.
The separately described learning module performs partial consistency inference using instantiated connections. It is not full arc/path consistency in every case, a mandatory part of every GDS run or a generic gradient-training procedure. Detailed later GENET mechanisms are outside the claims of this record.
The application and benchmark discussion does not establish universal success, polynomial expected time or exact equivalence to min-conflicts. No author implementation, learning procedure or benchmark was executed for this draft.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	broad_match	The inspected AIO target is a broader class of connected artificial-neuron units. This specific main-plus-guard architecture uses weighted inputs, binary threshold outputs and coupled neural units to encode a finite CSP. The target does not specify the guard asymmetry or row update. The mapping is not exact identity with HopfieldNetwork, whose pinned definition requires symmetric connectivity and content-addressable memory.

Evidence Sources & Primary Literature

2 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`johnston1989-gds`	Learning in Stochastic Neural Networks for Constraint Satisfaction Problems Mark D. Johnston, Hans-Martin Adorf	primary paper full_text	ID: https://ntrs.nasa.gov/citations/19900019762 sha256:f17944a5502c…
`minton1992-nasa`	Minimizing Conflicts: A Heuristic Repair Method for Constraint-Satisfaction and Scheduling Problems Steven Minton, Mark D. Johnston, Andrew B. Philips, Philip Laird	author technical report full_text	ID: https://ntrs.nasa.gov/citations/19930006097 sha256:4548b61515c4…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-15T21:19:50Z

Updated: 2026-10-15T21:19:50Z

Source File: data/curated/GuardedDiscreteStochasticNetwork.yaml