Geometry Dash RL

Geometry Dash is difficult for an AI agent because a level can rapidly switch between movement modes with completely different controls and physics. Geometry Dash RL builds a reinforcement learning agent that learns to play the game using gameplay state and repeated attempts, adapting its behavior across modes like cube, ship, and spider. The project aims to complete levels in as few attempts as possible, focusing on sample-efficient learning rather than simply memorizing a level.

Learning-Based Cache Eviction

Caches are only useful if they can quickly decide what data to keep and what to evict, but traditional policies like LRU struggle when access patterns change. Learning-Based Cache Eviction explores whether lightweight machine learning models can make better eviction decisions by learning from past access behavior. The project builds a cache simulator in Rust and compares learned policies against traditional heuristics across synthetic and real-world workloads.