My thesis work was applying a cutting edge AI algorithm to a board game called Ticket to Ride. The algorithm, AlphaZero, was arguably the best Chess and Go A.I. ever created. Given that it did not use historically collected professional moves to train, I figured this would be a great chance to apply it to a game I love that is lesser known.

The diagram above depicts the neural network portion of the algorithm, which serves to provide evaluations of given game states. Instead of directly choosing a move, the neural network is tasked with providing two outputs, (1) a winning probability v and (2) the probability distribution p across all possible moves depicting their chances of being played.

Given our model, we can do a Monte Carlo tree search which is guided by the neural network. As shown in the diagram above, the nodes are updated using the blue highlighted p and v values from the neural network, which are then used in selecting the nodes in later searches. The selection algorithm is as follows:

The idea, given the algorithm as described above, is to simulate games and do tree searches at every turn. The tree searches allow us to find real moves that result in winning. At the root node of the tree, move probabilities can be derived that will then train and align the network.

The description of the thesis above is a general synopsis. The thesis is a comprehensive explanation of the whole process. I learned so much from this thesis and it was one of my favorite projects to work on!