2023 ASEE Annual Conference & Exposition

In nature, organisms evolve into their own niches in an environment over time, despite harsh changes in both biology and nature itself. This paper describes the development and observations from a self-learning sandbox intended to mirror and emulate the biological systems around natural selection and environmental pressures. Unique pixel creatures with random behaviors are generated in a pre-made environment and compete against others to survive and pass their genetic information to the next generation. This iteration of self-learning differs from standard neural networks by the method of which fitness is achieved. As opposed to the model of backpropagation, which applies changes to errors after the learning action has been done on the same model, this instead uses a generalized fitness approach where only the top performers of each generation may be given the chance to move on. Random changes called “mutations” will give a varied approach, act partially in place of a learning rate, and prevent a form of the local minima problem, as well as provide resilience to environmental change or change from possible competitors. We develop a simulator based on an emulation modular framework with a customizable environment and varying levels of complexity. This sandbox encodes genetic data and abstracts the concepts of behavior and genotypes using machine learning concepts. Besides inputs and outputs, organisms’ internal networks are completely dependent on its encoded “genes”, a bit string, which includes connections between neurons and the properties of the neurons themselves. We develop such a sandbox, make conclusions and comparisons to nature, and give insight to possible expansion. We also evaluate the changes in configurations and its effects on unique trials within the simulator and advise how similar projects may proceed against some problems in design and theory.