Deep learning in computer chess
This report demonstrates a statistic chess position evaluation of a chess engine using two differ- ent neural network architectures. In this project, I worked on two deep learning techniques to achieve the aim of in a seven-class classification task using 10,000 distinct chess positions from Kaggl...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
Nanyang Technological University
2023
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/172008 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | This report demonstrates a statistic chess position evaluation of a chess engine using two differ-
ent neural network architectures. In this project, I worked on two deep learning techniques to achieve
the aim of in a seven-class classification task using 10,000 distinct chess positions from Kaggle1, la-
beled by Stockfish 16, a simple MLP and a three-layer neural network inspired by Mathew Lai.The
first implementation and architecture is a Multi-Layer Perceptron (MLP) which contains three layers of
128 neurons each with ReLU activation function and an output layer with CrossEntropyLoss function.
The second architecture is following the concept of Matthew Lai’s Giraffe to build a three-layer neural
network, whose first layer was not fully conneccted but seperated into three modalities piece-centri,
square-centric and position-centric to prevent overfitting, more details would be discussed in Method-
ology section. Eventually, the MLP achieved the test accuracy 84.26% and the three-layer Neural
Network achieved test accuracy 95.38% . |
|---|