- Basic idea is that robots are becoming more common, generating large amounts of data
- But the data is often poor; cheap, restricted sensors
- Argues that humans are good at making alot from a small amount of information (Hieder and Simmel)
- Deals a bit with nonstationarity where intents of agents can change in the middle of some interaction (the rules of who avoids who, chases who, in tag, for example)
- Experiments very similar to what was done in my cog sci paper
- Defines nonstationarity as a rapid change in mathematics of the estimated policy when it is represented as a polynomial
- Talks about finding the rules from tag, and catch, but didnt yet discuss how the inference is performed or what the hypothesis space was
- Learns faster when the machine can interact with the environment by testing hypothesis (active learning)
- Work on humans training robots when the humans perception system is limited to that of the robots
- This allows instruction to work much better than if the human has a richer set of data to work from
- Integration of heterogeneous sensing data from multiple robots