We consider a class of sequential environments that provide an agent
with propositional state descriptions as well as access to facts and
objects that may be necessary for carrying out optimal actions.
Learning the "state'' portion of these environments is related to
gathering experience in sequential decision problems, as considered in
the field of reinforcement learning. However, learning about the
schema underlying the facts and objects is more in the vein of concept
learning, as considered in the field of Inductive Logic Programming
(ILP). Despite the rich literature on theoretical efficiency in both
of these problems, algorithms focused on learning
in environments that combine them have produced largely empirical
results. We present an extended review of work in these areas, and
then provide a formalism for a restricted class of environments with
states, facts, and objects (ESFOs) that facilitates efficient
learning. We provide some example domains and present some
introductory results on the sample complexity of learning in
such environments. Finally, we discuss extensions of the formalism
and analysis that are of interest in a continued investigation of
these topics.