This paper is perhaps one of the most beautiful academic papers I’ve ever read, both in content and in aesthetics. The entire premise of the paper is that, using the set of skills available to an agent to construct symbolic representations of the agent’s domain is the best way to go about developing representations with which the agent uses to plan and perform domain tasks. Abstract representations developed in this manner can be used in varying complexity of state spaces, be it a grid world or the real world. This is simply genius.
I also love how George delicately crafts his sentences; patiently forming and shaping them into the masterpiece that it is. He writes so beautifully.
Here is a link to the paper: http://irl.cs.brown.edu/pubs/orig_sym_jair.pdf