Integrating Teaching and Reinforcement Learning in the Options Framework

I introduce a new way of building initial knowledge into a reinforcement learning agent. The agent will acquire a set of initial options by generalizing from training examples provided by a human supervisor. I show how this can significantly increase convergence speed compared to learning with primitive actions only. Moreover, using this method we can create a single option for dealing with similar situations. This means less memory for storing the option’s policy and less options to decide between at each step compared to creating one option for each particular situation.