Learning General and Efficient Representations of Novel Games Through Interactive Instruction

The goal of our research is to develop agents that can learn new tasks through real-time natural interactions with a human. Previously, we have described Rosie, an agent implemented in Soar that interactively learns new tasks. In this paper, we describe novel extensions to Rosie that allow it to learn complex, hierarchically defined concepts and to dynamically compile the interpretation of the task instructions. We evaluate the generality and efficiency of the knowledge that the agent learns over seventeen simple games and puzzles embedded in a variety of simulated and real world robotic environments. Our results show that learning hierarchical concepts allows Rosie to transfer learned knowledge to new tasks and decrease future instruction, while dynamic compilation decreases the time to process instructions, execute newly learned tasks, and learn future tasks.