OptiCarVis: Improving Automated Vehicle Functionality Visualizations Using Bayesian Optimization to Enhance User Experience

Research Background and Issues

• Identified Problems or Challenges:

  1. Current visualization designs for autonomous vehicle functionalities are often "one-size-fits-all" solutions, failing to accommodate the diverse perceptions and needs of passengers.
  2. Users exhibit significant variability in subjective evaluations of safety, trust, and acceptance, making it difficult for existing designs to balance multiple objectives within the design space.
  3. Traditional approaches, such as user-centered design and expert-defined standard designs, often require substantial resources for experimentation and fail to adequately capture individual preferences.

• Significance: The acceptance of autonomous driving technology is considered to depend on users' understanding of vehicle operations, detection, and planning functionalities. Enhancing user trust, perceived safety, and reducing cognitive load are critical factors in promoting public acceptance of autonomous driving.

• Research Motivation and Related Work:

  1. Previous studies have demonstrated that functional visualization can improve drivers' trust and reduce cognitive load (e.g., visualizations of semantic segmentation, internal function prediction, and trajectory planning).
  2. However, existing methods often rely on predefined designs or simple personalization, without fully leveraging optimization algorithms to explore broader design spaces.
  3. Multi-Objective Bayesian Optimization (MOBO) has shown high efficiency and performance in other design domains (e.g., touch keyboards, image classification), but its potential in the visualization design of autonomous vehicle functionalities remains underexplored.

Solution

• Method or Solution: The authors propose a method called OptiCarVis, which combines "Human-in-the-Loop" (HITL) interaction with Multi-Objective Bayesian Optimization (MOBO) to optimize the visualization design of autonomous vehicle functionalities within a larger design space, meeting individual users' subjective needs.

• Innovations:

  1. Utilizing the MOBO method to generate Pareto-optimal designs, balancing multiple objectives such as trust, safety, acceptance, aesthetics, and cognitive load without bias.
  2. Enhancing optimization efficiency and accelerating convergence through a Warm-Start HITL method that integrates expert-designed and user-customized design data.
  3. Incorporating end-users (including non-design-expert users) into the closed-loop design process, significantly reducing the resources required for traditional iterative design processes.

• Implementation Steps and Key Techniques:

  1. Define objective functions: Incorporate multi-level subjective user metrics (e.g., trust, predictability, perceived safety) into optimization objectives while considering the impact of design parameter mappings on cognitive load.
  2. Initialize design conditions: Include expert standard designs, user-customized designs, and randomly generated Cold-Start initial parameters.
  3. Implement MOBO using PyTorch's BoTorch library, optimizing design parameters with continuous value settings.
  4. HITL closed-loop optimization process: Iteratively present design variants and collect user feedback to optimize parameters until satisfactory designs are achieved.
  5. Conduct final experiments through online user studies to compare the advantages and disadvantages of designs under multiple conditions.

Research Outcomes

• Specific Results:

  1. The HITL MOBO method (especially its Cold-Start and Warm-Start variants) significantly improved users' subjective evaluations of safety, trust, predictability, and aesthetic characteristics.
  2. The MOBO method substantially reduced cognitive load and demonstrated higher efficiency in multi-objective optimization compared to non-MOBO methods (e.g., expert designs or user-customized designs).
  3. Users reported that the ability to personalize designs increased their satisfaction and engagement with the design process.

• Comparison with Existing Solutions and Advantages:

  1. Compared to Expert Designs: OptiCarVis's Pareto-optimal solutions addressed a broader range of user needs while effectively reducing the resources required for design.
  2. Compared to User-Customized Designs: MOBO-optimized designs avoided potential user misjudgments, such as cognitive overload caused by excessive information.
  3. Superior to Traditional A/B Testing: Demonstrated significantly improved efficiency in addressing complex multi-objective optimization problems.

• Experimental or Evaluation Results:

  1. In an online experiment involving 117 participants, Cold-Start and Warm-Start variants showed no significant differences but both outperformed non-MOBO methods, particularly in enhancing perceived safety and trust.
  2. Users expressed strong approval of the visual adjustments during the HITL design process, though improvements in aesthetics and acceptance were limited.
  3. Pareto frontier analysis identified optimal design parameter ranges, which closely aligned with user priorities (e.g., key visual focus areas).

• Limitations and Future Directions:

  1. The current study is limited to small-scale simulated driving scenarios and does not fully address the needs of complex driving situations.
  2. The use of explicit optimization (requiring users to provide frequent active feedback) may lead to user fatigue; future work should explore optimization methods based on implicit signals (e.g., eye-tracking, heart rate data).
  3. The impact of cultural differences on optimized designs requires further investigation, such as customization for users from different countries.
  4. Although the experiments were conducted in a screen-based simulation environment, they cannot fully replicate real-world driving scenarios; future research should validate the method's practicality with real vehicle experiments.

Conclusion

OptiCarVis provides an innovative, scientific, and efficient framework for optimizing the user visualization interface of autonomous vehicles within a larger design space, laying the foundation for personalized user experience design. This approach effectively enhances public acceptance of autonomous driving technology and holds potential for further improvements in user experience and design efficiency through implicit optimization in the future.