Marie desJardins, Department of Computer Science and Electrical Engineering
James MacGlashan, Department of Computer Science and Electrical Engineering
Intelligent agents are often given a set of abilities by a designer that the agent can use to accomplish various goals. However, if an agent is presented with a new complex problem that requires high-level abilities that a designer has not defined for it, then the agent may not be able to solve the problem in a reasonable amount of time. This makes agents very dependent on the designer to define an appropriate set of skills to use. Humans, in contrast, are born with very low-level abilities, and learn high-level reactive skills over time. In addition to learning these skills, humans can plan solutions to problems using the learned skills, allowing them to quickly find solutions to new problems. AI techniques, on the other hand, often use either a pure learning approach or a pure planning approach, which each have their own advantages and disadvantages.
Skill Bootstrapping is a novel integrated planning and learning framework that will allow an agent to start with primitive actions, and progressively learn generalized reactive skills that can be used during planning to quickly solve new, more complex, problems.