| dc.description.abstract | Lenia is a 2D cellular automata, with continuous states, space, and time, and has the ability to generate lifelike and captivating virtual creatures. Lenia has been shown to be a great way for experimenting with evolution and artificial life, as life in Lenia is highly flexible and graphic, and shown to be capable of movement, growth, consumption, reproduction, self-repair, self-defense, and swarming, among others. Although Lenia demonstrates a variety of fascinating behaviors, it is a highly complex system, making it incredibly time-consuming to explore the vast space of potential Lenia patterns. This is due to the presence of numerous patterns that bears similarities to one another within the system. When encountering a large solutions space, a known optimization technique is the use of quality diversity, which we aim to utilize. This approach entails seeking a diverse array of solutions and behaviors, each showcasing a unique approach to solving the task at hand. By doing so, we can effectively explore a broad range of possibilities and uncover a multitude of distinct strategies for addressing the given problem. We want to exploit this advantage of quality diversity, with the aim of discovering a wide range of patterns in Lenia. In order to obtain a diverse range of patterns, it is necessary to distinguish between the various behaviors exhibited by each pattern. Through the automation of behavioral descriptor creation, we use a neural network to capture the essential aspects of each pattern that may not be explicitly described manually, and differentiates the patterns for us. This thesis investigates whether automatically defined behavioral descriptors are helpful tools to achieve a wide range of patterns in Lenia. This was achieved by developing a new method for automatically searching through Lenia, by combining quality diversity and behavior descriptors. To better understand Lenia and achieve our goal, we focused on exploring pattern movement and used discovered patterns to pre-train behavioral descriptors. Using this method, we produced multiple collections of patterns with both animal and non-animal patterns. These collections were then used to analyze and evaluate the efficacy of our approach, by review of the state of the individuals, fitness, visualization of the patterns and archive, and inspecting the different groupings of patterns. This has enabled us to find strengths of our method, such as the method’s ability discover intricate patterns and differentiate between complex textures, allowing a higher interpretation of behavior descriptors. We believe that the presented method can provide new insight into searching through the space of Lenia. Furthermore, the use of automatic behavioral descriptors assists in understanding and analyzing digital life more effectively, and that further in-depth studies could reveal new potential patterns in Lenia. | eng |