Abstract
We present an agent, called RosieTAG, that is implemented in Soar and interacts with an external robotic environment. Rosie learns new games through interactive instruction with a human via restricted natural language. Instead of learning policy or strategy information, as is common in other game learners, Rosie learns multiple game formulations (the objects, players, and rules of a game) and then uses its own general strategies to solve them. We describe the structure and functionality of Rosie, and evaluate its competence, generality, communication efficiency, communication accessibility, and ability to continuously learn and accumulate new tasks and new task knowledge.
