GCAP 2019

Race for the Galaxy uses a temporal difference neural network to power its AI.  This knowledge-free system does not require human input to generate training data, which makes it extremely efficient for a small team with limited resources.  Even though our AI has trained on just over 30,000 games, because of this reinforcement learning model which uses predictive data to hone itself at the turn level, it has effectively benefited from over a million turn events.  The neural network we use actually has twelve bifurcations, each to support a specialized configuration of players, expansion content, and game preferences.  For each of these twelve, we support two flavors of neural network function, one to predict an opponent’s next move, and one to evaluate the board state, which is used when the AI simulates forward to choose its next move.  This talk will speak to how the Race for the Galaxy neural networks are architected, and how and when developers should use similar models.