Cost-Aware Bayesian Optimization for Interactive Devices

Paper Title

Cost-Aware Bayesian Optimization for Interactive Devices

Publication Info

• Topic area: Optimization techniques for iterative prototyping in interactive device design.
• Keywords: Bayesian optimization, cost-aware optimization, prototyping, interactive devices, iterative design, human-computer interaction, resource allocation, design space exploration, cost modeling, simulation studies.

Background and Problem

• Problem / challenge: Existing Bayesian optimization methods in HCI are cost-blind, treating all prototyping actions as equally expensive. This oversimplification fails in real-world scenarios where costs vary drastically between hardware and software changes, leading to inefficient resource allocation.
• Significance: Prototyping costs significantly influence what gets built, how often teams iterate, and which design directions are feasible. Addressing cost variability can improve efficiency and sustainability in design workflows.
• Motivation and related work: Prior HCI research has focused on reducing prototyping costs through toolkits, fabrication techniques, and computational feasibility assessments. However, these approaches do not address how cost influences the decision of what to prototype next. Bayesian optimization has been applied in HCI for design iteration but lacks cost-awareness, which is critical for real-world prototyping with heterogeneous costs.

Solution

• Proposed approach: The paper introduces a cost-aware extension to Bayesian optimization, named CABOP, which incorporates prototyping costs into the acquisition function to maximize expected improvement per unit cost.
• Novelty:
  1. Adapting cost-aware Bayesian optimization for interactive device prototyping by introducing a structured cost model.
  2. Defining a cost classification (tweak, swap, create) and leveraging a prototype record to track reusable components.
  3. Demonstrating the method's effectiveness through simulations and a user study, achieving comparable outcomes at significantly lower costs.
• Procedure and key techniques:
  1. Embed a cost model into the acquisition function to prioritize cost-effective sampling.
  2. Classify prototyping actions into three cost categories (tweak, swap, create) and assign numeric costs.
  3. Use a prototype record to track previously built components, enabling cost reductions through reuse.
  4. Implement a smooth relaxation of the cost model using RBF kernels for gradient-based optimization.
  5. Evaluate the method through simulations and a user study, comparing it to standard Bayesian optimization.

Results

• Concrete findings:
  • Simulations showed CABOP achieved comparable utility to standard Bayesian optimization at ≈70% of the cost.
  • Under strict budget constraints, CABOP outperformed the baseline threefold in terms of regret reduction.
  • In a user study with 12 participants, CABOP achieved equivalent joystick designs at ≈67% of the cost.
• Advantage over baselines: CABOP consistently reduced cumulative costs while maintaining or improving design quality. It dynamically adapted to cost asymmetries, modularity, and changing cost conditions, unlike the baseline.
• Experiments / evaluation:
  • Simulations: Tested on multiple benchmark functions (e.g., Rosenbrock, Ackley, Goldstein-Price, Levy) with varying cost structures, budgets, and complexities.
  • User study: Participants optimized joystick configurations (hardware and software) using CABOP and the baseline. Performance was measured through task utility and subjective ratings.
• Limitations and future work:
  • The current model assumes independent component costs, ignoring interdependencies.
  • It does not account for varying fidelity levels in prototyping.
  • Future directions include extending the cost model to handle multi-budget scenarios, integrating multi-fidelity optimization, and dynamically adapting acquisition strategies.

Summary

This paper presents CABOP, a cost-aware extension to Bayesian optimization tailored for prototyping interactive devices. By incorporating a structured cost model and leveraging a prototype record, CABOP prioritizes cost-effective design iterations. Simulations and a user study demonstrated that CABOP achieves comparable outcomes to standard methods at significantly lower costs, making it suitable for real-world design workflows. Future work aims to address interdependencies, fidelity variations, and multi-budget scenarios to further enhance its applicability.