Adaptive systems learn in dynamic environments by repeatedly sensing the world,
performing an action, and receiving feedback from the environment. The area of
reinforcement learning concerns agents that learn
sequential behaviors from experience; however, learning in complex domains is
excruciatingly slow. We are developing reinforcement learning methods that can
be guided
both by reinforcements provided by the environment
and abstract advice provided by a human teacher. In particular, we are
developing methods in which advice is given in ordinary natural language (which
is translated into formal advice using a
learned semantic
parser). By taking advantage of general advice on actions to perform in
certain situations, the agent's learning rate can be greatly accelerated. This
work is related to our work on
theory refinement
and
natural language learning.
Learning from natural-language advice and reinforcements is the topic of the PILLAR research project.