Agents that Learn from Human Teachers
Papers from the AAAI Spring Symposium
Andrea L. Thomaz, Chair
We believe that learning will be a key component to the successful application of intelligent agents in everyday human environments (physical and virtual). It will be impossible to give agents all of the knowledge and skills a priori that they will need to serve useful longterm roles in our dynamic world. The ability for everyday users, not experts, to adapt their behavior easily will be key to their success. Machine learning techniques have had much success over the years when applied to agents, but machine learnin techniques have not yet been specifically designed for learning from nonexpert users and current techniques are generally not suited for it out of the box.
The symposium covered a variety of topics. For example:
- How do everyday people approach the task of teaching autonomous agents?
- What mechanisms of human social learning will machine learning agents need?
- Are there machine learning algorithms that are more/less amenable to learning with nonexpert human teachers?
- What are proper evaluation metrics for social machine learning systems?
- What is the state of the art in human teachable systems?
- What are the grand challenges in building agents that learn from humans?