Factors Impacting Landscape Ruggedness in Control Problems: a Case Study

Saliby, Michel El; Medvet, Eric; Nadizar, Georgia; Salvato, Erica; Thomson, Sarah L.

Abstract

Understanding fitness landscapes in evolutionary robotics (ER) can provide valuable insights into the considered robotic problems as well as into the strategies found by evolutionary algorithms (EAs) to address them, ultimately guiding practitioners toward better design choices. However, most studies on fitness landscape analysis have been conducted on theoretical benchmarks, lacking direct relevance to practical robotics applications. This work aims to bridge this gap by (1) gathering a variety of measures to assess the ruggedness of a fitness landscape, (2) validating them on a simple benchmark problem with a known and tunable fitness landscape, and (3) applying these measures to a continuous control problem-a robotic navigation task. Using a highly customisable task, we investigate how various factors, including environmental conditions (i.e., the arena configuration), agent perception (i.e., the robot sensors), controller design (i.e., the structure of the artificial neural network controlling the robot), and fitness shaping (i.e., how the robot is rewarded for its behaviour) influence the ruggedness of the fitness landscape. Our findings suggest that simple measures can be sufficiently informative of the ruggedness of a given fitness landscape. Regarding the considered factors, we find that the ruggedness is primarily affected by the fitness shaping, followed by the controller features, while other factors tend to have a minor impact.

Citation

Saliby, M. E., Medvet, E., Nadizar, G., Salvato, E., & Thomson, S. L. (2024, September). Factors Impacting Landscape Ruggedness in Control Problems: a Case Study. Presented at WIVACE 2024 (XVIII International Workshop on Artificial Life and Evolutionary Computation), Namur, Belgium

• Presentation Conference Type: Conference Paper (published)
• Conference Name: WIVACE 2024 (XVIII International Workshop on Artificial Life and Evolutionary Computation)
• Start Date: Sep 11, 2024
• End Date: Sep 13, 2024
• Acceptance Date: Jun 30, 2024
• Deposit Date: Jul 5, 2024
• Peer Reviewed: Peer Reviewed
• Public URL: http://researchrepository.napier.ac.uk/Output/3703998