PROJECT TITLE :

Using Learning Classifier Systems to Learn Stochastic Decision Policies

ABSTRACT:

To solve reinforcement learning problems, several learning classifier systems (LCSs) are designed to be told state-action value functions through a compact set of maximally general and correct rules. Most of those systems focus primarily on learning deterministic policies by using a greedy action choice strategy. But, in observe, it may be more flexible and fascinating to learn stochastic policies, that can be thought-about as direct extensions of their deterministic counterparts. In this paper, we aim to achieve this goal by extending every rule with a brand new policy parameter. Meanwhile, a new method for adaptive learning of stochastic action selection strategies primarily based on a policy gradient framework has additionally been introduced. Using this technique, we tend to have developed 2 new learning systems, one based on a daily gradient learning technology and the opposite primarily based on a replacement natural gradient learning technique. Both learning systems are evaluated on three different varieties of reinforcement learning issues. The promising performance of the 2 systems clearly shows that LCSs provide a suitable platform for efficient and reliable learning of stochastic policies.