- Version
- Download 31
- File Size 2.45 MB
- File Count 1
- Create Date 20. October 2022
- Last Updated 20. October 2022
Bachelor Thesis
This thesis is written in german!
Title: Optimization of a chess engine’s evaluation function with efficiently updatable neural networks using the C++ library mlpack
Abstract: Efficiently Updatable Neural Networks (NNUE) have made an impact on the world of chess programs over the last years. They efficiently combine an evaluation function computed by an artificial neural network with the alpha-beta-search. The NNUE architecture is reconstructed using the C++ library mlpack, to be integrated in the chess program Heureka Engine. To achieve that, new network layers are designed and implemented. A training routine is developed to help supervise and evaluate the training. To be able to use the network during search, a representation of the network and efficient calculation of the forward pass are added to the engine. Finally, the results of the so far failed training experiments are presented and analysed with regards to the hyperparameters and optimizers used. The efficiency of the connection to the engine is also evaluated by considering further optimizations. By determining clear goals, the work on training the NNUE and on the engine may be continued. Parts of the implemented classes will be considered for inclusion in the used libraries.