Teaching a Ship (model) to Learn: From Classical Rules to Reinforcement Learning

TL;DR: Phase 1 gave us a classical navigation system that works. Phase 2 asked a harder question: can a neural network learn to navigate without being told the rules? Short answer: yes, sort of. It’s faster. It’s sometimes smarter. It also runs aground in ways a trained officer never would. Here’s what six months of RL training taught me about autonomous ships, and about the limits of learning from scratch.