EU4-Haven by Anthony

EU4-Haven by Anthony - October 2026 Scholarship Essay

The research question that changed how I study Computer Science was simple and stubborn: can a learning agent play a complex strategy game well enough to teach me what the game's hidden rules actually are? That question became EU4-Haven, a project where I train a PyTorch agent for Europa Universalis IV using behavioral cloning and reinforcement learning, while reverse-engineering the closed-source C++ engine that produces the observations the agent needs.

I started from curiosity, not from a clean dataset. The game exposes a huge action space and a dense world state. My early hypothesis was that if I copied expert play closely enough, the agent would generalize. Behavioral cloning on more than 6,600 expert trajectories got the policy moving, but it also revealed the gap between imitation and understanding. The agent repeated surface patterns and failed when the masked action set or the 247-dimensional observation vector shifted. That failure was the experiment. It forced me to ask better questions about representation, credit assignment, and what "knowing" a system means when you cannot read every line of the original engine.

To answer those questions I had to go under the hood. I mapped struct offsets, annotated thousands of functions in Ghidra, and built a DLL instrumentation layer with IAT proxy loading, vectored exception handling, and lock-free shared-memory IPC. Suddenly the research was not only about neural nets. It was about systems: how state moves across process boundaries, how a closed binary becomes a measurable environment, and how fragile assumptions break when timing and memory layout disagree with your model. Training on an NSF-funded HPC cluster made the scientific loop concrete. Change one assumption, rerun, compare. Curiosity without measurement is just preference.

The project deepened my understanding of Computer Science as a whole, not only machine learning. Algorithms class taught me complexity on paper. EU4-Haven taught me complexity under constraint: dynamic action masks, partial observability, and engineering debt that can invalidate a beautiful loss curve. The same habits transferred to Raider Marketplace, the campus platform I cofounded. There the research question was human rather than simulated: what product changes reduce friction for students who do not already know whom to ask? We shipped React/Vite and SwiftUI clients to more than 100 active users, then revised when listings stalled or trust broke. Hypothesis, release, measure. Different domain, same scientific muscle.

As a first-generation QuestBridge Scholar and Hispanic/Latino student at Colgate, raised by a single mother in Montgomery, Texas, I did not inherit a private research lab. Projects like EU4-Haven are how I build one. Studying CS at Yonsei this fall with Gilman STEM support stretches the same curiosity across a new academic culture. I still start with a question I can test, and I still refuse to call a result understanding until I can name the next experiment that would prove me wrong.