Neural network trained to play Mordhau.
Presentation: https://youtu.be/hWH5SrUWL9A Demo: https://youtu.be/Uv6v9SmMq18 (content warning: video game violence)
Shows the performance of the AI after training on about 24 minutes of gameplay data. More training data is expected to increase performance.
The neural network is an LSTM that learns to play like a human through supervised learning. Data is collected through a game process interface designed in C++ which is used to train a neural network in Python using Keras.
The neural network is given the game state which consists of the bone positions of the enemy (e.g. head, foot, hand) and is trained on output provided by a human. All behaviors exhibited by the program are learned and there are no pre-defined actions.
The process interface program utilizes Windows functions to read the process memory to extract the necessary data for the neural network. The interface utilizes an SDK generated by Unreal Finder Tool which extracts and provides a header-only library for reading from and writing to the necessary data structures of the game. https://github.com/CorrM/Unreal-Finder-Tool
The neural network is run within C++ using the frugally-deep library. https://github.com/Dobiasd/frugally-deep
Special thanks to Carlos Guerrero (https://github.com/chaquator) for help in designing the game process interface.