Variance and Distribution Models for Steering Tasks

Title of the Paper

Variance and Distribution Models for Steering Tasks

Paper Information

• Field of Study: Human-Computer Interaction (HCI), Behavioral Modeling
• Keywords: Steering law, movement time distribution, probabilistic modeling, HCI, steering tasks, trajectory modeling, skewed distribution, evaluation metrics, interaction design, uncertainty quantification

Research Background and Issues

Identified Problems or Challenges

1. The traditional Steering Law describes the linear relationship between movement time (MT) and the index of difficulty (ID), but it only models the mean of MT, failing to capture the variance or overall distribution of MT.
2. Providing only a point estimate of MT's mean neglects data dispersion and does not quantify predictive uncertainty.
3. There is limited research on skewed distributions (e.g., Gamma and Lognormal distributions) and their association with movement tasks. Specifically, Steering Law (similar to Fitts' Law in HCI task modeling) has not extended its analysis to variance and distribution-based insights.

Importance of the Research

1. Considering the variance and distribution of movement time allows for a more comprehensive understanding of human movement behavior.
2. Accurately predicting distribution parameters can better quantify task uncertainty, laying the foundation for more complex HCI modeling (e.g., nested menu selection).
3. Variance and distribution models can assist designers and researchers in optimizing interactive interfaces and devices.

Motivation and Related Work

• Steering Law is widely applied in HCI fields, such as virtual environment navigation and touch device evaluation, but existing models focus solely on mean prediction.
• Consistent with findings from Fitts' Law studies, evidence suggests that the standard deviation of movement time increases alongside the index of difficulty (ID).
• This study aims to propose a variance model for movement time based on ID and explore distribution modeling for movement time.

Solution

Proposed Methods or Solutions

1. Quadratic Variance Model:

   • Proposes a quadratic relationship between variance and ID: [ \sigma^2 = c + d \cdot ID^2 ] where (c) and (d) are empirically determined parameters.
   • This model assumes that movement comprises multiple sub-movements, each with a random duration.

2. Movement Time Distribution Modeling:

   • By combining the quadratic variance model with the Steering Law, the mean and variance of MT can be predicted.
   • Six probability distribution models were tested, including Normal distribution, truncated Gaussian distribution, Lognormal distribution, Gamma distribution, Extreme Value distribution, and ExpGaussian distribution.
   • Parameters were fitted using Bayesian posterior inference methods (Stan language).

Innovations

1. From Point Estimation to Distribution and Variance Prediction:
   • Overcomes the limitation of solely predicting the mean, providing multidimensional insights.
2. Introduction of Non-Normal Skewed Distributions:
   • Demonstrates that skewed distributions like Gamma and Lognormal better fit MT data.
3. Model Validation and Application Across Various Task Types:
   • Comprehensive experiments comparing six variance models and 31 distribution models.

Implementation Steps

1. Extract and preprocess experimental data, including touchpad data provided by Zhou and Ren, as well as mouse control experiment data collected by the authors.
2. Fit the mean and variance of movement time against ID.
3. Use distribution modules incorporating variance predictions to fit MT distributions, followed by model selection and evaluation.

Research Outcomes

Specific Results

1. Variance Modeling:

   • The quadratic variance model explained up to 94% of MT variance changes.
   • The model outperformed other candidates (e.g., linear or constant models).

2. Distribution Modeling:

   • Gamma, Lognormal, and Extreme Value distributions outperformed Normal and truncated Gaussian distributions in fitting MT distributions.
   • Variance and distribution modeling accurately predicted PDFs and CDFs, aligning well with real-world data.

3. Application Analysis:

   • Provides tools for quantifying specific event probabilities or task uncertainty based on MT distributions.

Comparison with Existing Solutions

• Compared to Steering Law models based solely on mean prediction, the proposed approach quantifies distribution characteristics, enhancing the explanatory power for movement behavior.
• Across diverse task environments (e.g., straight tunnels, curved tunnels), variance and distribution models maintained consistent predictive performance.

Limitations and Future Directions

1. Current Model Limitations:

   • The model's ability to explain variance decreases as ID increases.
   • Potential additional factors influencing MT variance remain unexplored.

2. Future Research Directions:

   • Extend existing models to capture variance and distribution characteristics in complex scenarios.
   • Investigate broader movement tasks and dynamic user behaviors, such as complex paths or multimodal interaction tasks.

Conclusion

This paper introduces a quadratic variance model and its integration with distribution prediction, contributing new methodologies to the Steering Law modeling domain. It expands the understanding of human movement behavior and provides theoretical support for optimizing complex interaction designs.