A major focus of my work was developing a reinforcement learning sensor tasking framework for space object detection. I constructed a Deep Q-Network (DQN) algorithm with convolutional architectures, experience replay, and reward shaping to train an agent to prioritize high value targets within stochastic environments (modeled via Gaussian Mixture distributions to resemble the random space domain). To stabilize training under sparse and delayed rewards, I implemented replay buffer management, target network synchronization, and exploration decay to mitigate divergence in these large environments. This system was designed to reduce uncertainty in the detection and tracking of space objects.

To enable quick experimentation, I implemented parallel processing (CUDA) for environment generation. I trained the model on terabytes of synthetic datasets using Git LFS and TensorBoard workflows to manage reproducibility and performance tracking.