Article,
Hierarchical Reinforcement Learning with Grammar-Directed GA-P
International Journal of Soft Computing, 1 (1): 52--60 (March 2006)
Abstract
We propose a grammatical approach to hierarchical
reinforcement learning. It is based on the grammatical
description of a problem, a complex task, or objective.
The use of a grammar to control the learning
process,constraining the structure of the solutions
generated with standard GP, permits the inclusion of
knowledge about the problem in a straightforward
manner, if this knowledge exists. When the problem to
be solved involves the use of fuzzy concepts,the
membership functions can be evolved simultaneously
within the learning process using the advantages of the
GA-P paradigm. Additionally,the inclusion of penalty
factors in the evaluation function allows us to try to
bias the search toward solutions that are optimal in
safety or economical terms,not only taking into account
control matters. We tested this approach with a real
problem, obtaining three different control policies as
a consequence of the different fitness functions
employed. So,we conclude that the manipulation of
fitness function and the use of a grammar to introduce
as much knowledge as possible into the search process
are useful tools when applying evolutionary techniques
in industrial environments. The modified fitness
functions and genetic operators are also discussed.
