DOOM game engine used to compute neural networks with "voodoo dolls"

Neural Networks in DOOM

The "Doomformer" project allows neural network computations to be performed directly within the DOOM game engine. This is achieved by mapping network values and operations to in-game elements, specifically referred to as "voodoo dolls".

How it Works

Each value in the neural network is represented by a "voodoo doll" within the DOOM map. The marine character can "ink the floor" to input data, and then "press READ" to initiate the computation. For an MNIST network, 1,355 voodoo dolls run the network, and the doll corresponding to the winning digit crosses a line to light up a display.

The project includes a Python interface for setup and verification. Python reads back the computed values from DOOM and checks them against NumPy's integer reference, ensuring exact matches with zero tolerance. Users can either have Python control the marine's actions or manually draw digits within the game.

Interactive Demonstrations

A "museum.wad" file is available, allowing users to play the demonstration without Python. In this mode, stepping on seven floor pads lights a digit, and pressing "READ" triggers a smaller network of 136 voodoo dolls to display the answer. The system also supports dynamic rewiring of connections, demonstrated by a "MUTATE" function that can change a recognized digit.

Technical Verification

The project emphasizes the exactness of the computations, stating that DOOM's computed values precisely match the NumPy integer reference. This verification process is compatible with various DOOM ports, including GZDoom, LZDoom, UZDoom, and ZDoom 2.8.1.