Adaptive Neural Network Usage in Computer Go

For decades, computer scientists have worked to develop an artificial intelligence for the game of Go intelligent enough to beat skilled human players. In 2016, Google accomplished just that with their program, AlphaGo. AlphaGo was a huge leap forward in artificial intelligence, but required quite a lot of computational power to run. The goal of our project was to take some of the techniques that make AlphaGo so powerful, and integrate them with a less resource intensive artificial intelligence. Specifically, we expanded on the work of last year’s MQP of integrating a neural network into an existing Go AI, Pachi. We rigorously tested the resultant program’s performance. We also used SPSA training to determine an adaptive value function so as to make the best use of the neural network