Analysis of rule learning with brain-likenetworks

Rule learning is a challenge for both artificial and biological learning systems. Whilethere are system that can learn complex tasks with rules, it is often impossible to understand what they have actually learned. In this thesis we have defined understandablenavigation tasks with clearly defined rules in a maze and in a binary tree environment. We have built a reinforcement learning agent with a shallow semi linear neural network that we test on these tasks. This agent is simple enough to analyse its inner workings and has the advantages of neural nets, like the potential for scalability and generalisation.