Date of Award

2023

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Joshua Bongard

Second Advisor

Dryver Huston

Abstract

The Baldwin effect is an evolutionary theory regarding the assimilation of ontogenetic changes into a population's genome via selection pressure to entrench beneficial phenotypes discovered through learning. In evolutionary computation, the incorporation of learning into non-embodied agents allows them to navigate otherwise rough fitness landscapes by allowing for local exploration at particular points in that landscape. Prior work investigating the specific mechanisms by which learned behavior is genetically assimilated is almost entirely limited to non-situated, non-embodied simulations such as bitstring manipulation. However, recent research has demonstrated that genetic assimilation can be observed in embodied agents. Learning more about the ways embodiment may affect the mechanisms of genetic assimilation can help us better understand how learning can affect evolution and enhance our design of evolved, learning, embodied systems.

To accomplish this, we co-evolve the initial values and learning rules for each synapse in the controlling neural network of three different robots to investigate the impact of morphology on the Baldwin Effect. We found that the different morphologies tested were capable of genetic assimilation within the evolutionary timespans provided and that each morphology exhibited significant differences in the amounts of genetic assimilation they were capable of. These differences were due entirely to different behaviors between morphologies, while the rate of genetic assimilation due to evolution of non-learning synaptic weights was constant regardless of morphology.

Language

en

Number of Pages

43 p.

Download

Share

COinS