Motives and Emotions in a General Learning System

TL;DR

This paper explores the crucial, yet often overlooked, role of motives and emotions in designing intelligent learning systems. The authors propose a framework within their BAIRN system where these internal states are not just emergent properties but integral components that guide learning, shape behavior, and contribute to the development of distinct "personalities" in AI.

What Problem Does This Paper Address?

The design of intelligent systems, particularly those capable of learning and adapting, has largely focused on cognitive processes and learning algorithms, often neglecting the influence of internal states like motives and emotions. This paper argues that for systems to be truly well-adapted and exhibit complex, individualistic behavior, we need to understand and integrate these motivational and emotional processes. The authors highlight fundamental questions: What is the adaptive function of motives and emotions in intelligent systems? What aspects are "innate" versus learned? And how do they influence learning, and vice-versa? They aim to provide a preliminary exploration of these issues.

Methods

The paper introduces the BAIRN (a general learning system) as a framework for exploring these concepts. BAIRN is a self-modifying information processing system that uses production systems as its core units. Key features relevant to this paper include:

  • World Model: A long-term memory (LTM) world model integrating procedural and declarative knowledge, represented as a network of nodes storing production systems.
  • Consciousness Levels: A three-level consciousness model (focal, peripheral, unconscious) managing parallel processing.
  • Prepotency Values: Nodes have "prepotency values" that determine their strength in bidding for operation, allowing for exploration of motivational structure's impact.
  • Time Line/Episodic Memory: A sequential record of system activity used for regularity detection and learning.
  • Motive Generators: These bias the system's world model construction and constrain behavior by scanning the time line for relevant experiences and initiating LTM modifications.
  • Emotion Generators: These act as "hard-wired" information sources that aid in discrimination and canalizing behavior. They link sensori-perceptual conditions with physiological actions, creating "innate" primary emotions (pleasure, fear, anger).
  • Adaptive Function: Motives are seen as essential for directionality and constraining action choices. Emotions provide information for discrimination and behavior canalization.
  • Learning Integration: The system learns new emotions and motives by processing sequences of primary emotion activations or by combining existing motive generators. Prepotency values are assigned to new nodes based on the strength of the involved motive generators and emotional intensity.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Motives as Directional Constraints: Motives are not just about "wanting" something, but about actively biasing the system's internal representation and behavior, providing essential directionality for flexible systems.
  • Emotions as Information Filters: Emotions serve as crucial "hard-wired" information sources, helping the system rapidly discriminate and canalize its behavior based on physiological and situational cues.
  • The Interplay of Arousal: General autonomic arousal levels influence the spread of activation in the LTM network, with high arousal leading to less constrained spread and low arousal to more focused attention.
  • Learning New Emotional and Motivational States: The system can learn more complex emotions and motives by observing patterns in its experience (time line), suggesting a pathway for developing sophisticated internal states.
  • Emergent Personalities: By manipulating initial repertoires of emotion/motive generators and learning parameters, BAIRN can develop distinct "personalities," offering a path towards self-modifying systems with individual characteristics.
  • Pathology through Extremes: The paper suggests that combinations of generalization/discrimination parameters that emphasize extremes could lead to emotions and motives analogous to human pathology.

Key Results

The paper presents a conceptual framework and a system design rather than quantitative results from extensive experiments. The "results" are primarily the demonstration of how BAIRN can:

  • Integrate motive and emotion generators into its learning processes.
  • Use these generators to influence the construction and modification of its world model.
  • Assign prepotency values to new knowledge based on motivational and emotional context.
  • Provide a mechanism for learning new, more complex emotional and motivational states.
  • Lay the groundwork for developing systems with emergent "personalities."

Limitations and Future Work

  • Preliminary Exploration: The authors acknowledge this is a preliminary exploration of complex issues.
  • Conceptual Framework: The paper primarily outlines a theoretical framework and system architecture rather than presenting extensive empirical validation.
  • Future Work: The authors suggest that manipulating initial repertoires and learning parameters can lead to emergent personalities. They also note that the approach could be applied to control third-generation robotic devices that need to function autonomously for extended periods. Further research could explore the specific combinations of parameters that lead to different types of "personalities" and investigate the implications of extreme generalization/discrimination settings.

Summary updated

BibTeX
@inproceedings{1983motives,
  title = {{Motives and Emotions in a General Learning System}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure