AIRIS Full

Home / Hyperon AI Algorithms / AIRIS / AIRIS Full

Responsible: Cook & Hammer 2024 introduced AIRIS; the paper-cited reference repository is Berick Cook's AIRIS_Public. Cook is at SingularityNET (berick.cook@singularitynet.io); Hammer is at KTH Royal Institute of Technology.

Papers: (KTH/DiVA portal record diva2:1890142, source spec); Goertzel (2025-09), Integrating SubRep with AIRIS (and PLN): General Direction and Minecraft Use-Case (Hyperon-side extension / boundary source — not the AIRIS source spec; see V1-2 below).

Status: [IMPLEMENTATION-BACKED-CORE] per Non-clustered HAA cluster pilot Source 1 reconciliation (closed 2026-05-06). The Cook & Hammer 2024 paper's core architecture (Rule Base + State Graph + Pre-/Post-Action Observation, with online rule induction and low-confidence-state curiosity) is implemented in berickcook/AIRIS_Public (Cook's reference / stable builds, 162 commits, last commit 2026-03-26) and in singnet/AIRIS-general (Cook's SingularityNET-side generalized rewrite, MIT, last commit 2026-03-13). Adjacent integration claims from the Goertzel 2025-09 paper (six-tuple rule structure, tripartite state, Monte Carlo rollouts) are paper-extensions and remain outside the implementation; this status describes only the Cook & Hammer 2024 core.

This card provides technical depth beyond the concise AIRIS index card. AIRIS (Autonomous Intelligent Reinforcement Inferred Symbolism) is a causality-based learning agent that builds expert-system-like causal rules from observations of changes in its environment caused by its own actions. Unlike Pearl-style causal networks, AIRIS does not require a pre-specified causal graph; unlike Cyc and NARS, it does not require human-encoded prior logical knowledge; unlike Markov Decision Process RL, it constructs future predicted states it has not previously experienced via rule application on a self-generated State Graph. The system retains expert-system advantages (transparency, scrutability, mutability) while supporting continuous online learning.

Related cards: ECAN Full (boundary — see V1-3), PLN Full (boundary — see V1-4), AtomSpace Full (boundary — see V1-5), Self-Modification and Safety Full (conceptual cross-link only — see "Self-modification scope" below), NARS Ecosystem (Hammer's primary work is NARS), MetaMo Full (Hyperon-side adjacency via Goertzel 2025-09 SubRep paper).

Core Concepts

Causal Rules. A Rule's Conditions consist of (1) the action that was taken, (2) the Original Value (the value prior to the action), (3) the Relative Change Positions (positions of other values that also changed, relative to Original Value's position), (4) the Relative Change Original Values (those values prior to the action), and (5) a list of other Rules observed to be true for the prediction. The Statement is what the Original Value changed into. (Cook & Hammer 2024, paper p.2 §2 "Rule Creation"; reference implementation airis_stable.py at create_rule L1280.) The Cook & Hammer formalism does NOT include a separately-formalized firing probability or motive-feature shift — those are Goertzel 2025-09 SubRep extensions (V1-2).

State Graph. AIRIS's internal world model. Rules are applied to the current State to predict modified States, which are stored in the State Graph. Planning traces an action path from current State to a Goal State that satisfies the agent's Goal Sub-States. The State Graph is self-generated by rule application — possible states are not pre-specified — and "mirrors the style and usefulness of finite-state machines" (paper p.3). The State Graph is NOT AtomSpace-backed in any audited implementation (V1-5); any AtomSpace integration is forward-looking design work, not a current implementation property.

State Confidence and Curiosity. Each prediction is annotated with a State Confidence ∈ [0,1] computed from the highest-scoring Rules' precondition match scores. When goal sub-states cannot be reached, AIRIS designates the State with the lowest Confidence as the new Goal State. Cook & Hammer explicitly call this "an effective model of artificial curiosity, which does not only capture aspects of information gain as in [Still & Precup 2012], but leads to deliberate interactions planned by the agent" (paper p.3). Reference implementation: airis_stable.py confidence_heap mechanism at L437–465 + L559. This is a self-contained planning/exploration heuristic, NOT an ECAN-style attention-allocation system (V1-3).

Self-modification scope. AIRIS continuously modifies its own Rule Base and State Graph via the Rule Creation and Rule Updating processes (paper p.5; airis_stable.py update_rule L1585). It is a self-LEARNING system. It does NOT modify its own source code, planner architecture, or rule representation; it is NOT a self-MODIFYING-CODE system in the AGI-safety sense. Cross-link to Self-Modification and Safety Full is conceptual only — the two concepts differ.

Comparisons. AIRIS is positioned in the Cook & Hammer 2024 paper relative to Pearl's causal networks (no pre-specified causal model needed), Cyc (no prior logical knowledge needed), NARS (no logical input encoding needed), PDDL planners (causal rule induction rather than hand-written domain knowledge), information-theoretic curiosity (Still & Precup 2012 — extends beyond information gain), and standard Reinforcement Learning (Sutton & Barto 2018, Moerland et al. 2023 — comparison baselines).

Implementations

• berickcook/AIRIS_Public — Cook's reference / stable builds. The implementation directly cited by the paper (Cook & Hammer 2024 footnote 3, p.2). Python; primary file airis_stable.py (1881 lines); puzzle-game and gym (cartpole, mountaincar) drivers; MNIST driver; "Minds Eye" planning-visualization tool. AGPL-3.0. HEAD f41041328d215ed623e08e015d6d88ddd2ee9d9a (2026-03-26). 37 stars, 12 forks.
• singnet/AIRIS-general — Cook's SingularityNET-side generalized rewrite (active 2026-03-13). Pure-Python single-file library (airis.py, 1417 lines) accepting structured (s_input, L24) and unstructured (u_input, L30) JSON-encoded data. Designed as a library drop, not a packaged application. MIT. HEAD 37861c953c338f53128e202ba49b4b0510c4bece (2026-03-13). Status: rewrite-WIP relative to AIRIS_Public.
• singnet/AIRIS-client — Minecraft client wrapper, NOT the AIRIS algorithm. Community-maintained (last HEAD by Ilya Fedotov, not Cook). HEAD 87238d28b10ccfd2137e6729c50f4857a38aba43 (2024-11-05). Status: client-only.

Wiki attribution corrections (V1-1). Until the 2026-05-06 cluster-close, the wiki's Reference+GitHub Repositories card was misattributed to singnet/AIRIS_Public (HTTP 404, does not exist) instead of berickcook/AIRIS_Public (the paper-cited canonical repo). Additionally, the AIRIS index card's +content referenced singnet/AIRIS-scripts which also returns HTTP 404. Both were corrected at cluster close. Per CF5.6.E ≥2-instance criterion (this V1-1 + the audit-batch oversight from 2026-05-05), the standing protocol "wiki edits referencing external URLs must include fresh WebFetch verification before the edit" is promoted from candidate to live.

Empirical Results (Cook & Hammer 2024 §5)

• Grid World Puzzle Game (13 levels of varying difficulty): untrained agent completes all levels in 4,612 timesteps in a single continuous episode (no episodic resets). Trained replay completes in 921 timesteps. A novel level not seen in training is solved in 67 timesteps (best possible 54). Demonstrates objective-changing without retraining and one-shot generalization.
• Cartpole (Geva & Sitte 1993 benchmark): AIRIS reaches first 200-score at 32 episodes and consistent solution at 130 episodes; DQN baseline first reaches 200 at 60+ episodes and consistent solution at 200+ episodes.
• MNIST: 91% accuracy on first training-set run, 97% on second; 90%→93% test set first run, 99% second. Comparison: deep CNN required several hundred runs to reach equivalent accuracy.

Limitations (Cook & Hammer 2024 §6)

AIRIS cannot learn rules for events that occur regardless of the agent's action, restricting it to environments where only agent-caused events occur. High-dimensional inputs (e.g., raw video) pose computational cost limitations. The paper notes "AIRIS will likely require additional attention mechanisms to prioritize processing of values relevant to the current task" — this is the forward bridge to ECAN-style attention allocation (cross-cluster boundary). Continuous-space and 3D extensions, plus Robot Operating System / Semantic SLAM integration for embodied operation, are listed as future work.

Boundaries

• vs ECAN (V1-3): AIRIS curiosity = lowest-State-Confidence-as-Goal-State, a self-contained local planning/exploration heuristic. NOT a reimplementation of ECAN's STI/LTI economic-attention allocation over AtomSpace. Cook & Hammer §6 explicitly notes AIRIS itself does NOT contain an attention-allocation system equivalent to ECAN's.
• vs PLN (V1-4): AIRIS State Confidence and rule-match scores are prediction-quality / planning heuristics, NOT PLN strength/confidence truth values (which are statistical-model truth values defined by PLN cluster Sources 9–11).
• vs AtomSpace (V1-5): AIRIS has its own State Graph world-model. The audited repos show no AtomSpace / MeTTa / OpenCog runtime binding; any AtomSpace integration is future design work, NOT current implementation.
• vs Self-Modification and Safety: AIRIS is self-LEARNING (online rule induction), NOT self-MODIFYING-CODE. Cross-link is conceptual only.
• vs SubRep / MetaMo (V1-2): Goertzel 2025-09 Integrating SubRep with AIRIS is a Hyperon-side design paper that extends AIRIS — its six-tuple rule, tripartite state, and Monte Carlo rollout language are NOT in Cook & Hammer 2024 or the audited repos. Treat as boundary/extension source.

Cluster-pilot extraction archive

Source 1 of the Non-clustered Hyperon AI Algorithms cluster pilot (closed 2026-05-06; further sources S2..S5 covering MetaMo, Pattern Mining, Concept Blending + Semantic Parsing, NACE + AI-DSL + MOSES + MeTTa-NARS pending). On-disk archive at scripts/archive/non_clustered_haa_pilot/source1_airis_paper/ with brief.txt + findings_codex.txt + findings_gemini.txt + findings_reconciled_crossmodel.txt. Local paper at publication_texts/2024_Cook_Hammer_AIRIS.pdf with provenance file alongside.