TLDR: This paper proposes a computational concept of the psyche, viewing it as an operating system for intelligent agents (human or AI) that manages life activities by satisfying needs. It introduces a “needs space” and uses emotions as reinforcement signals for learning. Intelligence is framed as a decision-making system that optimizes for goal achievement, risk minimization, and energy efficiency, using a “survival energy” metric. The model extends reinforcement learning with prospect theory and was experimentally validated in a “Self Pong” game, showing that negative reinforcement can hinder learning. The framework aims to formalize general artificial intelligence based on optimizing needs satisfaction in complex environments.
The quest to understand and replicate the human mind has long fascinated scientists. A recent research paper, “Computational Concept of the Psyche,” by A.G. Kolonin and V.G. Kryukov, proposes a novel framework for modeling the human psyche, viewing it as an operating system for living or artificial subjects. This concept aims to pave the way for building artificial general intelligence (AGI) by formalizing the psyche’s functions into a computable model.
The authors suggest that the psyche acts as an operational system, managing an individual’s life activities. Within this system, intelligence serves as the decision-making component, guiding actions to satisfy needs in response to external stimuli. This perspective integrates various disciplines, including psychology, systems analysis, and even microeconomics, to create a comprehensive model.
A central idea is that human intelligence, and by extension, artificial intelligence, continuously makes optimal decisions throughout its lifespan. The “optimality” of these decisions is defined within a system of needs, which are influenced by the environment and the physical organization of the intelligent agent. The paper emphasizes that the psyche unifies an agent’s thought processes, internal motivations, and its perception of reality.
The Role of Needs and Emotions
The research highlights the critical role of “needs” in driving behavior. These needs are not static; they are dynamic, competitive, and interdependent, as suggested by Maslow’s hierarchy and Anokhin’s work. The paper introduces a “needs space” or “needs matrix” as a core element of the model. This matrix describes the current state of the system by tracking the priority and satisfaction levels of various physiological and psychological needs. These needs can be hierarchical, ranging from basic individual survival to socio-economic and even civilizational levels.
Emotions are presented as crucial feedback mechanisms. They signal the significance of events in relation to an agent’s needs. Positive emotions arise when needs are met or goals achieved, while negative emotions indicate unmet needs, prompting the system to prioritize their satisfaction. This makes emotions a powerful reinforcement function for learning, with the intensity of an emotion directly correlating to the reinforcement value.
Decision Making and Survival Energy
Intelligence, in this framework, is the ability to achieve complex goals under challenging conditions with limited resources. The paper posits that decision-making is a multi-parametric optimization problem. The system constantly seeks the most effective solutions to satisfy current needs and anticipate future ones, balancing resource consumption and acquisition. This involves a continuous process of motivation, where actions are driven by emotions and internal states (prioritized needs, available energy) and external stimuli.
A unique contribution is the concept of “survival energy.” This is not just physical energy but a “universal currency” or “token” used to quantitatively assess physiological and psychological processes within the computational model. It allows for a standardized comparison of diverse needs and actions, aligning with economic principles where the psyche operates on principles of cost-benefit analysis, risk management, and energy efficiency. The goal is to maximize success in achieving goals, minimize existential risks, and maximize energy efficiency.
Learning and Computational Architecture
Learning in this model occurs through the acquisition of experience, which informs future decision-making. This experience, structured within the needs space, helps the system choose the most effective actions in terms of cost and benefit. The paper extends traditional reinforcement learning by incorporating “prospect theory,” which considers both positive and negative outcomes and their subjective utilities, rather than just expected utility. This leads to a more nuanced, multi-dimensional (vector or tensor) evaluation of utility, reflecting the complexity of real-world goals and motivations.
The proposed anthropocentric conceptual architecture of the psyche integrates this needs space with a decision-making system. It formalizes complex goals as a “motivational vector,” which is a product of long-term need importance (genetic/cultural code) and short-term need actualization (emotions). This mathematical model allows for the calculation of the expected value or utility of actions, considering various factors like energy efficiency and predictability.
Experimental Validation
To demonstrate the feasibility of their concept, the authors conducted a minimal experimental implementation. They extended a model for an agent playing “Self Pong” (tennis against a wall). The agent operated in a 4-dimensional needs space: positive reinforcement (Happy), avoidance of negative reinforcement (Sad), novelty of states (Novelty), and predictability of situations (Expectedness). The experiment explored different prioritization strategies for positive and negative reinforcement.
A key finding was the negative impact of punishment on learning. When the priorities for positive reinforcement and avoiding negative reinforcement were equal, learning slowed down or even became impossible. This was attributed to the suppression of the agent’s exploratory activity due to penalties for errors made while testing new strategies. This suggests that the balance between reward and punishment is crucial for effective learning in intelligent systems.
Also Read:
- Reinforcement Learning: The Core Driver for Advanced AI Research Systems
- Optimizing Industrial Scheduling: A Novel DRL Environment for Flexible Job-Shops
Towards General Artificial Intelligence
Ultimately, the research posits that human intelligence constantly evaluates whether a decision will lead to a better or worse outcome. The psyche manages the satisfaction of needs, both physical and psychological, within an individual’s activities. By quantifying the degree of need satisfaction through energy expenditure and time, this model provides a framework for understanding the core task of intelligence.
The ability of an intelligent agent to operate in complex environments, considering multiple goals and threats within its individual space of biological, physical, and existential needs, and optimizing their satisfaction, can be seen as the definition of general intelligence. The strength of this intelligence is determined by the dimensionality and complexity of this needs space. The computational realization of this cognitive architecture offers a promising path toward creating advanced artificial intelligence systems. You can read the full paper here.
