Morphological development, exploration, and solution guidance on the fitness landscape

What is it about?

In human beings, the morphological development of the body that takes place in parallel with the development of the cognitive system has been shown to facilitate the acquisition of new skills and abilities. In the literature, applying these natural principles to robotics has yielded mixed results, finding that the joint development of the morphology and cognitive system may favor learning, being irrelevant and even detrimental. Also, in the literature, one of the most popular hypotheses states that morphological development improves learning thanks to increasing the exploration of the solution space, avoiding stagnation in local optima thanks to continuous changes of the morphology, which implies a continuous variation of the morphology and control relationship. In this article, we studied the influence of a single morphological development strategy, growth, and its nuances as a technique to improve the exploration of the solution space and, thus, favoring learning. With this goal in mind, we performed a series of experiments over two different robotic morphologies, a quadruped and a bipedal one, in the task of learning to walk, where the results of learning while growing and learning while some noise is added to the morphology or control system are compared. The comparison is done with the aim to identify whether growth-based morphological development is just another optimization technique that only favors the exploration of the solution space, or if it has some nuances that make it different. To analyze this hypothesis, we have made use of the Search Trajectory Network representation, to visualize and analyze the evolution of the genotypic space of each type of experiment during learning, which also allows us to compare its evolution. Finally, the results indicate that noise and growth increase exploration and may yield results that increase the learning performance, but only growth consistently guides the learning algorithm toward the area of optimal solutions.

Featured Image

Why is it important?

This work is relevant for the following reasons: - It shows how there is a relationship between the morphology and the solution space (or fitness landscape in most cases), where it can be observed how progressive transitions from one morphology to another, also implies a progressive transition from one solution space to another. - Morphological development is not simply another optimization technique to favor exploration and avoid stagnation in local optima. It has other nuances, related to the development of the morphology that implies guidance of the learning algorithm towards the areas of optimal solutions.

Perspectives