Leveraging Human Insights by Combining Multi-Objective Optimization with Interactive Evolution.

摘要

Deceptive fitness landscapes are a growing concern for evolutionary computation. Recent work has shown that combining human insights with short-term evolution has a synergistic effect that accelerates the discovery of solutions. While humans provide rich insights, they fatigue easily. Previous work reduced the number of human evaluations by evolving a diverse set of candidates via intermittent searches for novelty. While successful at evolving solutions for a deceptive maze domain, this approach lacks the ability to measure what the human evaluator identi es as important. The key insight here is that multi-objective evolutionary algorithms foster diversity, serving as a surrogate for novelty, while measuring user preferences. This approach, called Pareto Optimality-Assisted Interactive Evolutionary Computation (POA-IEC), allows users to identify candidates that they feel are promising. Experimental results reveal that POA-IEC finds solutions in fewer evaluations than previous approaches, and that the non-dominated set is significantly more novel than the dominated set. In this way, POA-IEC simultaneously leverages human insights while quantifying their preferences.