The goal of this project was to demonstrate the flexibility of shumai, the exciting new ML Library for Bun JS runtime. In an ideal world, we'd create a model that would trade on technicals through neuroevolution, but crypto is hardly a rational market, so YMMV. Highly recommend only using this in sandboxed env (default) and not actually connecting to a Coinbase account.
The initial population is built given a specified network topology and assigned random weights. For each generation, these weights are randomly mutated and each network is assigned a fitness value based on its performance. A pooling algorithm is then used to select models for the next generation and the process is repeated. Models with a higher fitness tend to make it to the next generation where their weights are mutated and ideally create a higher performing generation.
See Golang Implementation for the original Go implementation that inspired this project.
Next, install project dependencies:
bun installOnly macOS and Linux are supported. Linux installs default to GPU computation with CUDA, and macOS to CPU. Detailed install instructions here: shumai repo.
This demo project leverages Bun's FFI and Zig to execute native code helpers (specifically leveraging Zig's std.rand to randomly init weights/biases). Accordingly, you'll need to ensure Zig is installed in order to build the bindings; refer to the Zig installation guide for more details.
Once Zig is installed, you can build the bindings locally by running the following from the root directory of the project:
bun run gen:bindingsbun run start