Self-mediated Exploration in Artificial Intelligence Inspired by Cognitive Psychology

This research paper explores the integration of cognitive psychology principles into artificial intelligence (AI) systems to enhance their exploratory capabilities. The paper draws inspiration from cognitive psychology's extensive observations of the connections between epistemic and achievement emotional states and exploratory behavior in humans. The authors posit that by embedding these emotional states into AI learning methodologies, they can create agents that are more self-motivated to explore and acquire knowledge. The research argues that AI's current lack of exploratory capabilities hinders its autonomy and adaptability. Unlike even the simplest organisms, AI systems struggle to actively seek out new information and experiences, limiting their ability to learn and adapt to novel situations. The paper aims to address this deficiency by drawing upon the insights of cognitive psychology. The study's primary goal is to demonstrate that by explicitly modeling epistemic and achievement emotions within AI systems, researchers can induce exploratory behavior analogous to that observed in humans. The authors propose that by leveraging deep learning methodologies, they can computationally replicate the cognitive processes underlying these emotions and their influence on human behavior. The paper highlights the significance of this research for both AI development and our understanding of human cognition. By creating AI agents that exhibit emotion-driven exploration, the research contributes to building more autonomous and adaptive AI systems. Moreover, the study offers a unique perspective on how emotional states might drive knowledge acquisition and learning processes in humans.

The authors draw upon a substantial body of literature in cognitive psychology that highlights the link between epistemic and achievement emotions and exploratory behavior in humans. Studies cited in the paper demonstrate that experiencing surprise (epistemic emotion) when encountering information that contradicts existing knowledge leads to a desire to seek further information, driving exploratory behavior. Conversely, feelings of pride (achievement emotion) following successful task completion can induce a search for similar situations that evoke the same internal reaction, promoting further exploration. The paper contrasts the existing AI approaches to modeling emotional states with the proposed method. Current AI systems often define internal states arbitrarily to serve specific purposes, lacking a strong connection to psychological research. The authors argue that their approach, based on explicit modeling of epistemic and achievement emotions, provides a more realistic and psychologically grounded representation of emotional states.

The research employed a multi-layered neural network architecture to simulate the cognitive processes underlying emotion-mediated exploration in humans. This system is composed of three interconnected modules: 1. **Task-Oriented Module:** This module serves as the primary task-performing component of the AI system. It is trained using supervised learning techniques on a dataset of handwritten digits (MNIST). In the main experiment, this module receives new images for classification and generates predictions, which are subsequently used to trigger emotional responses. 2. **Actor Module:** This module is responsible for determining the exploratory rate of the system based on its current emotional state. It receives as input a single value representing the emotional score (either surprise or pride) and outputs an exploratory rate, indicating how much more data should be analyzed to mitigate the emotional state. 3. **Critic Module:** This module evaluates the actions (exploratory rate) taken by the actor, providing feedback based on whether the chosen rate improves the system's performance. It receives as input both the emotional score and the exploratory rate, and outputs a signal that reinforces the actor's decision-making based on its effectiveness in achieving the task objective. To simulate the influence of surprise and pride on exploratory behavior, the researchers developed separate functions for each emotion. These functions were based on standard performance metrics, such as task accuracy and confidence. The surprise function was designed to produce higher values when the system encounters high-confidence errors, indicating unexpected results. Conversely, the pride function was designed to increase with increasing task accuracy, reflecting feelings of accomplishment. The research used a learning cycle, inspired by cognitive psychology experiments with human participants, to train the AI agents to associate specific emotional states with corresponding exploratory behaviors. This cycle involved multiple episodes, each consisting of 20 steps. At each step, the task-oriented module classified a randomly selected image, triggering an emotional response based on its prediction accuracy and confidence. The actor then chose an exploratory rate, and the critic evaluated the effectiveness of this choice based on the subsequent performance of the system. The researchers conducted two separate experiments, one for surprise and one for pride. In each experiment, 250 distinct AI agents were trained using the learning cycle. The researchers then assessed the relationship between the emotional state of the agents and their exploratory behavior, measuring the strength of the correlation over time.

The research findings demonstrated a strong correlation between surprise and exploratory behavior in the AI agents. As the agents experienced higher levels of surprise, they were more likely to engage in exploration, seeking out additional information to resolve the unexpected results. This finding aligns with the observed behavior in humans, where surprise often motivates individuals to seek out further clarification or knowledge. The study also found a weaker relationship between pride and exploratory behavior. While some agents showed a slight increase in exploratory rate with rising pride, the overall trend was not as pronounced as with surprise. This suggests that while pride can influence exploration, it is not as strong a driver as surprise. This finding also aligns with cognitive psychology research, which indicates that pride might have a weaker impact on knowledge exploration compared to other emotions. The researchers observed that over time, AI agents exhibited a significant decrease in surprise levels and a slight increase in pride levels. This suggests that the learning process effectively reduced surprise through exploration and increased pride through successful task completion. This pattern aligns with the free-energy principle, which posits that self-organizing systems tend to minimize uncertainty and maximize reward. The results were robust across multiple agents, demonstrating that the observed relationship between emotional states and exploration was not simply a product of random variation. Instead, the consistent pattern observed across multiple agents suggests that the AI systems learned to effectively associate emotional states with appropriate exploratory behaviors, similar to humans.

The research findings support the hypothesis that explicitly modeling epistemic and achievement emotions within AI systems can induce exploratory behavior analogous to that observed in humans. The strong correlation between surprise and exploration, as well as the weaker influence of pride, aligns with previous cognitive psychology research. This suggests that the proposed approach provides a valuable framework for bridging the gap between AI and human cognition. The study's findings highlight the potential of emotion-driven exploration for enhancing AI autonomy and adaptability. By equipping AI agents with the capacity to experience and respond to emotions, researchers can create systems that are more self-motivated to learn and adapt to changing environments. This has significant implications for a wide range of applications, from robotics and machine learning to social and industrial settings. Moreover, the research provides a new perspective on the role of emotional states in cognitive processes. The paper's findings suggest that emotions, particularly surprise, might be a more powerful driver of knowledge acquisition and learning than previously thought. This opens up new avenues for research on human cognition and behavior, potentially leading to a deeper understanding of how emotions influence learning and decision-making.

The research demonstrates the feasibility of integrating cognitive psychology principles into AI systems to induce emotion-driven exploration. The study successfully replicated previous cognitive psychology findings on the link between surprise and exploration, highlighting the potential of emotion-based approaches for enhancing AI autonomy and adaptability. The study's findings provide a valuable framework for further research on AI-driven emotion and cognition, paving the way for more self-motivated and adaptable AI systems.

The study's limitations include the reliance on a single task (image classification) and the simplified representation of emotional states. Future research could explore the generalizability of the findings to a wider range of tasks and more complex emotional models. Additionally, the study focused solely on the relationship between individual emotions and exploration, neglecting potential interactions between multiple emotions. Further research could investigate the interplay between different emotional states in influencing exploratory behavior.