[論文レビュー] Computational Concept of the Psyche

Anton Germanovich Kolonin, Vladimir Krykov|arXiv (Cornell University)|Mar 16, 2026

Psychiatry, Mental Health, Neuroscience被引用数 0

ひとこと要約

要約: 論文は人工汎用知能のための心の計算モデルを提案し、欲求・感覚・行動の空間で意思決定を整理し、ハイブリッド神経-symbolic アーキテクチャ内で最小の ping-pong エージェントを実証します。

ABSTRACT

This article presents an overview of approaches to modeling the human psyche in the context of constructing an artificial one. Based on this overview, a concept of cognitive architecture is proposed, in which the psyche is viewed as the operating system of a living or artificial subject, comprising a space of states, including the state of needs that determine the meaning of a subject's being in relation to stimuli from the external world, and intelligence as a decision-making system regarding actions in this world to satisfy these needs. Based on this concept, a computational formalization is proposed for creating artificial general intelligence systems for an agent through experiential learning in a state space that includes agent's needs, taking into account their biological or existential significance for the intelligent agent, along with agent's sensations and actions. Thus, the problem of constructing artificial general intelligence is formalized as a system for making optimal decisions in the space of specific agent needs under conditions of uncertainty, maximizing success in achieving goals, minimizing existential risks, and maximizing energy efficiency. A minimal experimental implementation of the model is presented.

研究の動機と目的

• Motivate the need for a systemic, interdisciplinary model of the psyche for artificial agents.
• Define an anthropocentric cognitive architecture where the psyche manages needs, sensations, and actions.
• Formalize decision-making with a multi-objective, energy-efficient utility framework incorporating uncertainty.
• Introduce a hybrid neuro-symbolic architecture enabling system 1/system 2 interaction in AI.
• Demonstrate a minimal experimental implementation to illustrate experiential learning in the proposed space of states.

提案手法

• Propose a space-of-states model with a dedicated space of needs (needs matrix) and a tensor-based representation for complex interdependencies.
• Formulate decision-making as multi-objective optimization using a prospected utility that combines long-term priority (x) and current need actualization (y) under uncertainty.
• Incorporate prospect theory and energy efficiency as core factors in reward/utility calculations, extending traditional reinforcement learning.
• Introduce a hybrid neuro-symbolic architecture linking neural/associative representations with symbolic hypergraphs or graphs to enable system 1/system 2-like reasoning and knowledge transfer.
• Describe a four-layer memory architecture (episodic memory, model memory, short-term memory, attention focus) for robust learning and memory management.

実験結果

リサーチクエスチョン

• RQ1 How can a space of needs and a tensor representation capture the motivational dynamics of an intelligent agent?
• RQ2 Can a multi-objective, energy-aware utility framework effectively guide decision making under uncertainty in both artificial and biological-inspired agents?
• RQ3 What is the role of a hybrid neuro-symbolic architecture in enabling efficient learning and knowledge transfer for AGI?
• RQ4 How does experiential learning in this psyche space facilitate the progression toward artificial general intelligence?

主な発見

• Present a conceptual cognitive architecture where the psyche functions as an operating system managing needs, sensations, and actions.
• Introduce a survival-energy based universal currency to compare physiological/psychological processes and guide goal-directed activity.
• Demonstrate a minimal ping-pong agent in a four-dimensional needs space, illustrating experiential learning and the influence of priority vectors on behavior.
• Propose a four-layer memory system ensuring long-term episodic coherence while enabling model-based and symbolic reasoning with memory constraints.
• Show how a hybrid neuro-symbolic framework supports bidirectional transfer between neural and symbolic representations to improve inference efficiency and lifelong learning