Computational Rationality as a Theory of Interaction

Title of the Paper

Computational Rationality as a Theory of Interaction

Paper Information

• Domain: Human-Computer Interaction and Cognitive Science
• Keywords: Cognitive modeling, computational rationality, interaction, reinforcement learning, adaptability, individual differences

Research Background and Problem

• Identified Problems or Challenges: How do people’s interaction behaviors with computers adapt to their cognitive limitations, device design, and environmental conditions?
  • Traditional cognitive architecture models require researchers to assume task completion strategies and pre-encode rules, a process that is complex and constrained by the diversity of user strategies.
  • Existing models struggle to predict how users choose strategies in new task environments or designs.
• Significance: Understanding adaptability in interaction is crucial for improving interface design, optimizing interactive systems, and enhancing the collaborative capabilities of AI systems.
• Research Motivation and Related Work:
  • Recently, a new theory—computational rationality—has emerged, focusing on explaining and predicting how users adapt to cognitive and environmental constraints, emphasizing the mechanisms of human interaction behavior adaptation.
  • Existing work has shown promising explanatory power in domains such as driving, text input, visual search, and multitasking.

Solution

• Proposed Approach:
  • The authors propose "computational rationality" as a theoretical framework, combining "cognitive architecture" from cognitive science with "bounded optimality" from machine learning.
  • They use partially observable Markov decision processes (POMDPs) to model task decision problems and employ reinforcement learning (RL) algorithms to generate optimal strategies.
  • The theory assumes that user behavior is an optimal response to subjective rewards and constraints.
• Innovations:
  • Replaces rule presets in traditional cognitive architectures with computational optimization to generate task-specific strategies.
  • The theory starts from the user's internal environment (cognitive state) rather than directly modeling the external environment, offering a new explanatory framework for human internal processing mechanisms.
• Implementation Steps and Techniques:
  • Define interaction tasks as decision problems (POMDPs) with core elements such as states, actions, transition functions, reward functions, and observations.
  • Model users' cognitive limitations (including time, noise, uncertainty, and capacity constraints) through observation functions and transition functions.
  • Use RL to approximate optimal strategies and validate their predictive capabilities.

Research Outcomes

• Specific Results:
  • Developed a unified theoretical framework for deriving computational models involving cognitive constraints.
  • Provided explanations for adaptive behaviors in various interaction tasks (e.g., text input, menu selection) that align with experimental data.
  • Built specific computational models to predict user behavior and design personalized interactive interfaces.
• Advantages Compared to Existing Solutions:
  • Eliminates the need for complex predefined operational rules, directly predicting strategies from task goals and constraints.
  • The model can predict behavioral differences arising from individual variations or cognitive limitations.
• Experimental or Evaluation Results:
  • Experiments show that the computational rationality model effectively predicts strategies in visual search tasks, driving behavior adjustments, text input speed, and error rates.
  • Enables optimization of diverse design solutions to accommodate individual user capabilities, such as designing optimized keyboard layouts for patients with tremors.
• Limitations and Future Directions:
  • The current model is limited to explaining micro-HCI interactions and needs expansion into macro-level aspects such as social interaction, motivation, and context.
  • Further research is needed on strategy inference mechanisms during human learning and adaptation processes, as well as methods for handling contextual influences in complex environments.
  • Accelerating RL optimization speeds to address larger design spaces.

Conclusion

Computational rationality offers a novel approach that deeply integrates cognitive science and machine learning to explain the adaptive characteristics of human interaction behavior. Although the research has not fully addressed macro-level interactions, it has made significant contributions to micro-level interaction design and cognitive modeling. It lays a solid foundation for future expansions into areas such as motivation, emotion, social interaction, and design optimization.