Coletta, a Johnson Scholar, double majored in computer science and music, and minored in creative writing. She presented her Honors Thesis on Evolutionary Control of Micro Aerial Vehicles in Simulation via Zoom on Wednesday April 14th.
This thesis examined the challenge of safely landing quadcopters in simulation, using OpenAI gym environments to evaluate various machine learning algorithms. The two main categories explored were Deep Reinforcement Learning and Evolutionary algorithms. While the success of the DRL approach motivated this work, the evolutionary angle is of more interest due to its roots in more realistic biological inspiration. One specific evolutionary approach, NEAT, had success in the three-dimensional version of the problem, while none of the DRL attempts were successful. The dominance of NEAT on this challenge, while impressive in contrast with DRL, also had reasonably comparable success to a heuristic, human engineered approach.