Code for the project in the DL4G module at HSLU. The goal of the project is to build a bot that can play the game of Jass. Various methods of game play were explored. Rule-based, MCTS, DMCTS, DNN-based. Our bot utilizes DMCTS and a DNN network trained on trump selection data.
To tackle Jass's imperfect information, the bot employs DMCTS with multiple determinizations processed in parallel across all CPU cores within a 30-second decision window.
The DNN used for trump selection is a six-layer dense network, incorporating relu activations, regularization via L2, and dropout layers to mitigate overfitting, culminating in a softmax layer for final decision output.