The Cross Entropy Method for Fast Policy Search

We present a learning framework for Markovian decision processes that is based on optimization in the policy space. Instead of using relatively slow gradient-based optimization algorithms, we use the fast Cross Entropy method. The suggested framework is described for several reward criteria and its effectiveness is demonstrated for a grid world navigation task and for an inventory control problem.

• Type: Conference
• Year: 2003
• Where: ICML
• Authors: Shie Mannor, Reuven Y. Rubinstein, Yohai Gat
• Keywords: Cross Entropy Method, ICML 2003, Machine Learning, Markovian Decision Processes, Slow Gradient-based Optimization