I built a custom Deep Learning framework in pure Rust just to simulate Arknights: Endfield gacha luck (Talos-XII)

The author developed Talos-XII, a project initially conceived for simulating Arknights: Endfield gacha pulls. This project evolved into a highly optimized system featuring a custom Deep Learning engine built entirely in Rust. The core objective is to employ Reinforcement Learning algorithms, specifically PPO and DQN, to discover optimal gacha pulling strategies for budget-conscious players. The engine avoids Python, utilizing pure Rust and featuring a custom reverse-mode Autograd system. Performance is enhanced through Rayon for parallel tensor operations and hand-written SIMD kernels for critical paths. The model incorporates a Deep Belief Network for environment noise simulation and a Transformer architecture for the agent. Optimization draws inspiration from the DeepSeek mHC paper, presenting an interesting implementation challenge. Talos-XII simulates millions of pulls to determine the probability of obtaining specific characters with free resources. The project, essentially a "Neural Luck Optimiser," aims to guide players on optimal resource saving. Currently, Talos-XII is a command-line interface only, with no graphical user interface developed yet. The project repository and a relevant reference paper are provided for those interested. The author specifically acknowledges the DeepSeek team's mHC paper for its significant influence on the optimizer's design.