Introduction
Interactive Evolutionary Computation (IEC) facilitates solving problems where the "best" solution is subjective, such as in art, music, or coffee blending. In standard IEC, a human acts as the fitness function, manually selecting the best candidates in each generation. However, scaling this to multiple users introduces two major hurdles: human fatigue, where users tire of evaluating hundreds of iterations, and the "too many cooks" problem, where conflicting preferences among multiple users hinder the evolutionary process.
The authors address these challenges by proposing Collaborative Interactive Evolution (CIE), a framework where multiple users provide subjective input at the start, and a Genetic Algorithm (GA) handles the resolution of conflicts and exploration of the fitness landscape.
The CIE Framework and Methodology
Unlike traditional IEC, which requires constant feedback, CIE gathers user preferences once at the beginning to define the search space. Users design characters through a web-based system, which then feeds into the evolutionary process.
To resolve conflicts between different users' inputs, the system utilizes Fitness Function Templates (FFTs):
- FFT1 (Frequency-based): Awards fitness based on how many users requested a specific trait, modeling a "most popular" approach.
- FFT2 (Diversity-based): Rewards chromosomes that match many of a user's requested traits or those that match traits of a user who might otherwise be ignored. This prevents a dominant subgroup from silencing minority preferences.
The GA runs for 100 generations using these templates to evolve candidates, which are then compared against "Functional" methods like simple averaging (mean) and frequency (mode).
Experimental Setup
The study generated a live dataset through two experimental phases involving 10 users. Users designed characters with 12 traits (6 physical, 6 psychological), resulting in a 36-bit chromosome.
- Phase 1: Consistent participation over five days.
- Phase 2: Intermittent participation over seven days, simulating real-world usage.
Key Results and Insights
The researchers evaluated success through user voting and engagement metrics.
User Preference and Engagement
In both phases, users preferred CIE-evolved characters over functional ones by at least 2-to-1. In Phase 2, FFT2 was the clear winner, garnering 42% of the votes. Engagement metrics also showed that users spent significantly more time examining CIE characters. For instance, in Phase 2, FFT1 characters were viewed for 9,423 seconds total, compared to just 3,403 seconds for functional characters.
Conflict Resolution and Minority Protection
The authors argue that the stochastic nature of GAs is a feature for conflict resolution. While averaging produces "middle-of-the-road" results, the GA explores a broad spectrum of unpredictable but high-quality characteristics. FFT2’s reward structure ensures that even if a small group has very different tastes, their "genetic material" survives into the final pool. Even when two users conspired to dominate the results, the GA framework limited their influence to about 75% of the genetic material, ensuring the final output remained a true group product.
Conclusion and Future Work
CIE successfully mitigates human fatigue by moving user interaction to the onset of evolution. It also provides a robust mechanism for group collaboration, even acting as a catalyst for spontaneous social dynamics outside the system. Future work includes testing the framework in Single Display Groupware (SDG) environments, using meta-evolution to allow groups to vote on the fitness templates themselves, and applying CIE to broader domains like architecture or collaborative music composition.
Summary updated