Cognitive Active Inference

This repository contains a validated discrete Active Inference package, a generalized-coordinate continuous agent, matrix utilities, knowledge-base tools, and reproducible validation commands.

Use Python 3.10 or newer:

python -m pip install -e ".[dev]"

The editable install exposes the cognitive, Things, and scripts Python packages and these commands:

cognitive-create-node --help
cognitive-verify-links . --json
cognitive-validate-docs . --json
cognitive-benchmark --repetitions 10
cognitive-build-manuscript --output build/manuscript

DiscreteGenerativeModel validates the five matrices used by the package:

• A[o, s] = P(o | s) and B[s_next, s_prev, a] = P(s_next | s_prev, a);
• C contains finite observation log-preferences;
• D and E are normalized state and action priors.

All dispatcher methods return finite normalized distributions. The dispatcher supports variational, mean_field, and sampling inference, discrete policy sequences, explicit risk, ambiguity, epistemic information gain, horizons, temperatures, and seeds.

expected_free_energy returns the canonical objective and its components:

G = D_KL(q(o) || p*(o))  +  E_q(s)[H(P(o | s))]
    \_____ risk _____/     \____ ambiguity ____/

Ambiguity equals H[q(o)] - I(s; o), so the expected information gain is already inside it; the epistemic gain is returned as a fourth value for inspection and is not subtracted from the total a second time. InferenceConfig exposes exploration_weight for callers who want an explicit additional information-seeking bias — it defaults to 0.0, and 1.0 reproduces the objective this package used before 1.1.0. See CHANGELOG.md.

import numpy as np

from cognitive import ActiveInferenceDispatcher, DiscreteGenerativeModel, InferenceConfig, ModelState

model = DiscreteGenerativeModel(
    A=np.array([[0.9, 0.1], [0.1, 0.9]]),
    B=np.stack([np.eye(2), np.array([[0.1, 0.9], [0.9, 0.1]])], axis=2),
    C=np.array([0.0, 1.0]),
    D=np.array([0.5, 0.5]),
    E=np.array([0.5, 0.5]),
)
dispatcher = ActiveInferenceDispatcher(
    InferenceConfig(
        method="variational",
        policy_type="discrete",
        temporal_horizon=2,
        learning_rate=0.5,
        precision_init=1.0,
        seed=7,
    ),
    model,
)
state = ModelState(model.D.copy(), model.E.copy(), 1.0, 0.0, 0.0)
beliefs = dispatcher.dispatch_belief_update(1, state)
policies = dispatcher.dispatch_policy_inference(state)
assert np.isclose(beliefs.sum(), 1.0)
assert np.isclose(policies.sum(), 1.0)

Things.Simple_POMDP.SimplePOMDP provides a validated discrete POMDP with seeded sampling, persistence, histories, expected-free-energy components, and temporary-directory-friendly plotting. Things.Continuous_Generic provides precision-weighted generalized-coordinate updates and Pillow-backed multi-frame GIF animation through ContinuousVisualizer.

cognitive.utils.create_node.NodeCreator resolves configured paths relative to its YAML file, renders templates safely, and rejects unsafe names. The network visualizer uses deterministic layouts and handles empty graphs.

python -m pytest -q --cov
ruff check .
mypy code/tools/src code/Things code/scripts
python -m compileall -q code
python code/scripts/validate_docs.py --json
python code/scripts/verify_links.py . --json
python code/scripts/check_markdown_links.py . --json

--cov measures branch coverage over every module in the configured source tree against the floor in pyproject.toml. The report carries no include filter, so the number is the package's, not a chosen subset's: docs/tools/coverage_tools.md.

The documentation validator enforces the term policy declared in docs/policy/documentation_terms.yaml.

The complete executable manuscript lives in docs/manuscript/. Its builder generates deterministic figures, auto-numbered equations and tables, Pandoc citations, HTML, and a XeLaTeX PDF from docs/manuscript/config.yaml. Pass --no-render to run every stage the package owns without Pandoc or XeLaTeX:

cognitive-build-manuscript --no-render --output build/manuscript

Tests write artifacts only to temporary directories. Generated reports and visualization trees are intentionally excluded from version control.

The conceptual material in knowledge_base/ explains the mathematics and domain context; executable behavior is defined by the package and its tests.