"Reducing state space exploration in reinforcement learning problems by rapid
identification of initial solutions and progressive improvement of them"

Abstract:

Reinforcement learning is mainly used when an agent needs to explore a
new environment and gradually find good solutions by trial-and-error.
It is also typical that the "goodness" of an action taken by the agent
is not known immediately, leading to the need to treat "delayed
reward".  It is often argued that reinforcement learning is the kind
of learning that resembles learning by humans and animals the most.

Maze route finding is a typical reinforcement learning problem, where
the success of an action might not be known before the goal has been
reached.  Classical reinforcement learning is based on propagating
this success backwards to all the action decisions having participated
in finding the route, which leads to slow learning and necessitates
numerous tries, often using Monte-Carlo simulation.

The neural network that is presented uses a new network structure
based on "short term memory" for treating the current problem to be
solved and "long- term memory" for storing old solutions to the
problem.  The reinforcement learning used is of one-shot type and does
not require defining real-valued reinforcement values, so despite its
simplicity, it should be a good candidate for many practical problems.

Remark: The presentation will include an on-line computer simulation.
The emphasis of the presentation will be on the big principles, not on
the theoretical (mathematical or neurological) proof.