Entry Date:
July 21, 2008

Learning in the Large

What if we want machines to learn to behave in complex environments with:
(*) thousands of state variables
(*) hundreds of competing objectives, and
(*) partial observability

Current learning techniques work well in carefully framed domains. We will try to dynamically construct appropriate abstractions of current state. The goal is to achieve higher fidelity for aspects of the state that are important, likely, and will happen soon.