Rock Paper Scissors

I have released a rock paper scissors (RPS) AI on my homepage. It was made using hierarchical temporal memory. RPS is a great domain to test temporal pattern recognition, so I am somewhat disappointed how poorly my AI does. Some people may be surprised that there is any strategy involved in RPS. After all, an AI which just outputs random moves will win 50% of the time. Humans are terrible at being random, so its not hard for an AI to win against people. When competing AI against AI, superior strategies can be determined based on their performance against a large number of opponents. This allows some AIs to get much better results than ones which just output random moves.

I have become really interested in hierarchical temporal memory since starting grad school. My supervisor Nando said that this might be a topic I can pursue for my thesis. I think there are some pretty major flaws with the implementation by Numenta. I implemented my own version using backpropagation neural networks. So far I have tested it in three applications (binary input prediction, handwritten digit recognition, and RPS). The results haven't been great. Next I plan on comparing it directly to NuPIC on their standard datasets. Hopefully my implementation does better! If so, this would be a great thesis topic.