Visualizing Hyperparameters in 2D Drone Navigation
Maanas Punuru, Jan Ole Ernst
Affiliation: Panther Creek High School
IJSCAR Vol. 3, Issue 2 (2026) · pp. 4–9
DOI: 10.67149/yhjs2024.5/bx8r3n6w
Abstract
Reinforcement Learning (RL) is a subfield of Machine Learning that involves agents learning what actions to take given the current state and an eventual goal. RL has become prevalent in the modern world. RL-powered self-driving drones are used in modern society for various tasks such as delivering packages. Thus it would be practical and interesting to design an RL algorithm that attempts to teach a drone how to reach a target. However small-scale learning is difficult due to the problem of learning to take sequential beneficial steps especially with stochastic wind. In this paper we address this issue by testing various hyperparameters in an attempt to mitigate these problems. These small-scale tests can be crucial to large-scale drone navigation problems helping improve contemporary algorithms. In this paper we test hyperparameters such as unique reward shaping formulae and discount factors. Multiple experiments conclude that some hyperparameters have a large effect on drone performance while some do not.
Keywords: Reinforcement Learning (RL), Unmanned Aerial Vehicle (UAV), Hyperparameters, Reward shaping, Discount factor, Learning rate, Actor-Critic, Stochasticity, Proximal Policy Optimization (PPO)