A State Space Compression Method Based on Multivariate Analysis for Reinforcement Learning in High-Dimensional Continuous State Spaces