23 lines
3.4 KiB
TeX
23 lines
3.4 KiB
TeX
% ---------------------------
|
||
\section{Related Work: Existing Taxonomies and Frameworks}
|
||
% ---------------------------
|
||
Standards bodies and policy groups have invested heavily in AI definitions, lifecycle models, and governance instruments. However, none provides a compact, mechanism-centric taxonomy spanning \Lrn, \Evo, \Sym, \Prb, \Sch, \Ctl, and \Swm, nor an explicit grammar for hybrids.
|
||
|
||
\paragraph{Standards and terminology.}
|
||
ISO/IEC 22989 standardizes terms and core concepts for AI across stakeholders, serving as a definitional foundation rather than a technique taxonomy. ISO/IEC 23053 offers a functional block view for \emph{machine-learning-based}~ AI systems (data, training, inference, monitoring), which is valuable architecturally but limited to ML and therefore excludes non-ML pillars such as symbolic reasoning, control/estimation, and swarm/evolutionary computation \citep{ISO22989,ISO23053}.
|
||
|
||
\paragraph{Risk and management frameworks.}
|
||
NIST’s AI Risk Management Framework (AI RMF 1.0) provides an implementation-agnostic process for managing AI risks (govern, map, measure, manage). Its companion \emph{AI Use Taxonomy}~ classifies human–AI task interactions and use patterns. Both are intentionally technique-agnostic: they can apply to any implementation class, but do not sort systems by operative mechanism \citep{NISTRMF,NISTAI2001}.
|
||
|
||
\paragraph{Policy classification tools.}
|
||
The OECD Framework for the Classification of AI Systems organizes systems along multi-dimensional policy axes (People \& Planet, Economic Context, Data \& Input, AI Model, Task \& Output). This is a powerful policy characterization instrument, yet it remains descriptive and multi-axis rather than a compact mechanism taxonomy with hybrid syntax \citep{OECDClass}.
|
||
|
||
\paragraph{Regulatory regimes.}
|
||
The EU Artificial Intelligence Act introduces risk-based classes (e.g., prohibited, high-risk, limited, minimal) and obligations, largely orthogonal to implementation specifics. Technique details matter for \emph{compliance evidence}, but the Act does not define a canonical implementation taxonomy \citep{EUAIAct}.
|
||
|
||
\paragraph{Academic precedents and surveys.}
|
||
The textbook tradition organizes AI by substantive pillars—search/planning, knowledge/logic, probabilistic reasoning, learning, and agents—closely aligning with the mechanism families in this paper but without proposing a stable naming code or formal hybrid grammar \citep{AIMA4}. Reinforcement learning texts formalize optimization and value iteration for \Lrn/\Sch~ couplings \citep{SuttonBarto2018}. Classical theory anchors \Prb~ (\citealp{KnillPouget2004}), \Ctl~ (\citealp{Kalman1960,Pontryagin1962,TodorovJordan2002}), and foundational dynamics for \Evo~ (\citealp{Price1970,TaylorJonker1978}). Learning rules for \Lrn~ include Hebbian and Oja’s formulations \citep{Hebb1949,Oja1982}, while resolution proofs formalize \Sym~ \citep{Robinson1965Resolution}. No-Free-Lunch results motivate preserving multiple mechanisms rather than collapsing them into a single “optimization” bucket \citep{Wolpert1997}.
|
||
|
||
\paragraph{Gap and contribution.}
|
||
Taken together, these works motivate \emph{two layers}: (i) policy/lifecycle/risk instruments that are technique-agnostic and (ii) a compact, biologically grounded \emph{implementation taxonomy}~ with explicit hybrid composition. OPT fills the second layer with seven frozen roots and a grammar for hybrids, designed to interface cleanly with the first layer.
|