CuMind: a MuZero reimplementation

CuMind is MuZero from scratch: a representation network that encodes observations, a dynamics network that predicts the next latent state and reward, and a prediction network for policy and value, tied together with Monte Carlo tree search for model-based planning. It runs on Gymnasium environments.

I built it to understand the algorithm, and most of what I learned was about systems. Model-based planning is limited by simulation throughput and tree-search parallelism at least as much as by the algorithm. The matrix computations and the data pipeline needed heavy tuning before the agent could train in reasonable time, and the tree search is the part that resists batching.

It is a side interest, not a research direction, but it shaped how I think about where the time goes in a training loop.